Automated computer vision and dose–response modeling improve throughput and accuracy of an ex vivo functional precision medicine platform Authors: Noah Bell, Andrew Buckley, Breanna Mann, Xiaopei Zhang, Adebimpe Adefolaju, Rajaneekar Dasari, Rami Darawsheh, David E. Kram, Shawn Hingtgen, Andrew B. Satterlee Published: 2025 ABSTRACT Functional Precision Medicine platforms face significant obstacles to clinical translation, including stringent requirements for analytical consistency. In this study, we present an automated data analysis workflow for an organotypic brain slice culture–based functional assay that addresses these concerns by combining computer vision and dose–response modeling approaches. Automation reduces analysis time by approximately 99% compared to the original manual workflow (from ~ 20 h to ~ 15 min to analyze an 11-drug assay), increasing throughput capabilities and freeing researchers from tedious analyses. Comparison of automated measurements with previously published manual results revealed that automation increased consistency both within experiments and across replicate experiments by removing human subjectivities. Our workflow demonstrates how to implement computer vision using limited computational resources and minimal programming expertise by leveraging cloud-based, user-friendly ML tools (e.g., Biodock) integrated into local scripts. It also addresses challenges in automated modeling of datasets that include diverse dose–response behaviors, such as hormesis and plateau effects. Finally, we outline a validation strategy for automated analysis in the absence of gold-standard measurements, an issue common in functional assays where manual measurements involve user subjectivity.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-31666-y. Introduction As biological and clinical research generates increasingly large and complex datasets, automation has become essential for efficient, consistent, and objective analysis. Advances in imaging technology have enabled the production of vast quantities of intricate imaging data; manual analysis of these images and downstream processing of the resultant data are often prohibitively time-consuming1. Moreover, manual image analysis is inherently subjective and often results in discrepancies among users2,3; dim signals or nuanced features may be difficult to reliably detect by eye4. The increasing demand for efficient and objective analysis has spurred the development of numerous automated analysis tools. CellProfiler5 and ImageJ6 offer modular tools for designing custom automated computer vision (CV) workflows across scientific domains, although they lack support for labeling and training instance segmentation machine learning (ML) models. For some domain-specific images, such as stained images of muscular cross-sections, researchers in the field have developed custom tools for those commonly analyzed image types1,3,7,8. However, if a unique platform lies in a gap unfilled by either general or specific automation solutions, researchers must address the limitations of manual image analysis by developing a custom workflow. Researchers may encounter challenges such as training an ML model9,10, validating automated measurements in the absence of a gold-standard dataset11,12, or modeling the resultant measurements13,14. Our group encountered the bottleneck of manual analysis while developing an ex vivo platform for studying brain tumor treatment. This platform uses organotypic brain slice cultures (OBSCs) as living tissue substrates that host and sustain both tumor cell lines as well as living, uncultured brain tumor tissues resected from UNC Hospitals15. Engrafting patient tumors within the OBSCs’ natural brain architecture preserves the tumors’ original genetic and phenotypic characteristics and improves tumor survival ex vivo. In addition to studying the genetic and phenotypic characteristics of engrafted tumors, we treat each tumor with panels of therapeutics and directly assess antitumor potency and selectivity. We summarize both tumor response and off-target toxicity to the OBSCs themselves using a multiparametric Drug Sensitivity Score (DSS). Calculating DSSs for a single tumor sample against a panel of therapeutics required approximately twenty hours of active analysis, including: (1) manually measuring a series of spatially dependent signals from various complex image types; (2) combining and processing the resulting measurements to generate dose–response data; and (3) calculating multiple parameters via linear interpolation from each set of dose–response curves to generate each DSS. This complexity significantly limited our throughput. To continue advancing the OBSC platform toward clinical validation, we needed an automated workflow that maximized analytical consistency and efficiency. Here, we present a fully automated data analysis workflow for an OBSC-based functional precision medicine platform that addresses many common obstacles to developing a customized workflow (Fig. 1). A user-friendly app ensures organized collection and storage of experimental metadata. A computer vision (CV) process, incorporating the machine learning (ML) web app Biodock via API16, performs multimodal image analysis that is tailored to the diverse quantitative image-based data produced by the OBSC platform. Automated measurements of bioluminescence-based tumor kill were validated by comparison with previously published manual measurements15, and discrepancies were further investigated to determine whether automation improved the results. A multistep approach to dose–response modeling (DRM) uses the resulting tumor and OBSC survival measurements to model treatment effects and extract relevant response parameters. Our data show that these models, combined with improved statistical characterizations of dose–response, more accurately capture tumor-specific efficacy in the resulting DSSs. Furthermore, calculating the DSS for a single tumor sample against a panel of eleven therapeutics now takes 15 minutes instead of 20 h. By increasing the efficiency and reliability of data analysis, this automated workflow advances the translational potential of the OBSC functional diagnostic platform. Fig. 1OBSC workflow and impact of automation. Top Panel: Co-culture of tumor tissue with organotypic brain slice cultures (OBSCs), drug treatment, and imaging. Manual analysis requires ~ 20 h per assay and involves subjective ROI placement. Bottom Panel: Automated workflow using machine learning–powered computer vision for segmentation, automated dose–response modeling, and Drug Sensitivity Score (DSS) calculation. This reduces analysis time to ~ 15 min per assay and improves reproducibility. OBSC workflow and impact of automation. Top Panel: Co-culture of tumor tissue with organotypic brain slice cultures (OBSCs), drug treatment, and imaging. Manual analysis requires ~ 20 h per assay and involves subjective ROI placement. Bottom Panel: Automated workflow using machine learning–powered computer vision for segmentation, automated dose–response modeling, and Drug Sensitivity Score (DSS) calculation. This reduces analysis time to ~ 15 min per assay and improves reproducibility. While open-source platforms such as ImageJ and CellProfiler offer powerful capabilities for image segmentation and analysis, our workflow addresses several limitations that arise when applying these tools to complex, multimodal datasets like those generated by the OBSC-based functional assay. First, conventional tools often require manual tuning of macros or plug-ins for each image type, which can be time-consuming and difficult to standardize across experiments. In contrast, our workflow integrates a machine learning model trained specifically on brightfield images that contain OBSCs, enabling robust segmentation even in the presence of challenging features such as overlapping tissues, reflections, and low signal-to-noise ratios. Second, our approach leverages API-based integration with Biodock, allowing seamless incorporation of ML-generated masks into downstream analysis workflows without requiring manual intervention. Third, the workflow is designed to handle multiple image modalities (fluorescence, bioluminescence, brightfield) and link them to structured metadata for dose–response modeling, which is not natively supported in general-purpose tools. Finally, our validation against previously published manual measurements and the development of a dynamic dose–response modeling framework further distinguish this workflow as a tailored solution for functional precision medicine applications, rather than a general image analysis tool. Results Overview of the OBSC assay To measure therapeutic efficacy against tumor–OBSC co-cultures, microtumors are seeded onto OBSCs and allowed to engraft for 24 h before treatment is applied (Supplemental Fig. 1 A). If microtumors are generated from a cell line that already expresses the mCherry fluorescent marker, we acquire Day 1 Tumor images (D1–T) at this time point to establish an initial, pre-treatment size for each engrafted microtumor. If microtumors are generated from patient tumor tissues, they are first rapidly infected with mCh/FLuc reporter genes via lentiviral transduction (which takes approximately three days for full expression), then thoroughly washed before engraftment onto OBSCs (Supplemental Fig. 1B). Twenty-four hours after engraftment, microtumors are exposed to six increasing concentrations of a therapeutic, and survival is assessed 72 h after treatment initiation. Day 4 Tumor images (D4–T) are then acquired to measure final microtumor survival by capturing the FLuc bioluminescent signal emitted by each surviving tumor cell. In a parallel experiment, OBSCs without tumors are exposed to the same treatment and doses for the same duration, and Day 4 OBSC images (D4–O) are acquired to measure treatment-induced toxicity using the fluorescent cell death marker Propidium Iodide (Supplemental Fig. 1 C). The data from these three image types (D1–T, D4–T, and D4–O) are combined to calculate the percent survival of each microtumor or OBSC after treatment, measured relative to the survival of untreated tumors or OBSCs. On-target, treatment-induced tumor kill is then compared to off-target toxicity to OBSCs across eleven dose–response parameters. These weighted parameters are combined to generate the single summarized DSS, where a score close to 100 indicates high tumor-specific killing and a score close to or below 0 indicates little or no tumor-specific killing. Metadata input form Automation of data analysis in the OBSC workflow required an organized system for storing and managing experimental metadata, ensuring that any generated results could be contextualized and easily accessed for future analysis. To ensure consistency and accuracy, we developed a simple application that enables users to generate correctly formatted metadata (Supplemental Fig. 2). Users can select tumor lines and treatments from an internal database that stores additional information about each tumor and treatment. After users define the experimental metadata and specify dose groups, the app creates a CSV file ready to be populated with data from the rest of the workflow. For each microtumor engrafted on an OBSC, 45 metadata fields are completed and stored in a remote database. 31 fields are separately collected for each OBSC in a second, distinct, assay that measures off-target toxicity of a therapeutic to the OBSCs themselves. As of 05 August 2025, we have collected and stored 15,681 unique experimental results from 13,280 microtumors and 2,401 OBSCs in this database, totaling 672,031 individual data points. This manuscript details how analysis was automated for 886 microtumors and 1723 OBSCs previously published in Cell Reports Medicine15. Automating image analysis using machine Learning-Powered computer vision Of the three image types previously introduced, two (D4–T and D4–O) share the same format: a multimodal image of a six-well plate featuring a signal image, measuring a biologically relevant marker, overlaid on a grayscale image that captures the visible OBSCs. A method to identify individual OBSCs in the grayscale image would facilitate further analysis for both image types. The grayscale images contained features that made conventional image analysis techniques challenging, such as bright reflections from the plates and OBSCs that were occasionally touching the well walls or each other. Therefore, we trained an ML model to identify grayscale OBSCs using the web-based platform Biodock16. Biodock accelerates the development of ML models for biological applications by providing pre-trained models and an AI-assisted image labeling workflow. We trained an ML model on Biodock using 110 images containing 1,037 OBSCs, with 99 images used to train the model and 11 images reserved for model validation (Supplemental Table 1). All training and validation images were sourced from a separate set of experiments distinct from those previously published in Cell Reports Medicine to avoid overfitting the model to experimental images analyzed in this study. The model achieved a Mean Average Precision over an Intersection over Union, IoU, range of 0.5–0.95 of 99.1 on the validation images (Supplemental Table 2), closely reproducing the ground truth results (Fig. 2A). This accurate identification of OBSCs in training images generalized well to unlabeled images from the experiments included in this paper (Fig. 2B). Biodock’s extensive API functionality allowed us to incorporate the generated OBSC masks into custom, locally executed workflows. The location of each mask within an image was used to determine from which well, and therefore which dose group, each OBSC belonged. D4–T images feature two bioluminescent microtumors per OBSC, one seeded on each hemisphere. To measure D4–T signals using the OBSC masks, each mask is bisected via a rotated bounding box, creating one mask per OBSC hemisphere. Then, within each hemisphere mask, the triangle thresholding algorithm17,18 identifies the tumor signal present (Fig. 2C). This approach can accurately quantify the sizes of two microtumors with vastly different signal magnitudes within the same image (Fig. 2D). Our approach also accounts for microtumors that have been completely killed and emit no bioluminescent signal; this was done by measuring the background noise in signal-free OBSC hemispheres. Similarly, Biodock-generated OBSC masks were also used to calculate OBSC toxicity from D4–O image data. The average PI-derived fluorescence within each mask was measured to quantify cell death within each OBSC (Fig. 2E). D1–T images were much simpler to analyze, each featuring an mCherry signal emitted by a single microtumor against a dark background. The triangle thresholding algorithm was again used to identify the mCherry signal. This algorithm was designed to differentiate dim signals from a large, dark background, and it performed well on D1–T images when tested across a variety of microtumor sizes and morphologies. Triangle thresholding even recognized tumor features that were obscured by computer display settings (Fig. 2F), demonstrating the ability of a relatively simple algorithm to identify signals more precisely and consistently than a human technician. Fig. 2Visualization of computer vision applied to various tumor signals. (A) Two of the labeled OBSCs withheld from the ML training process in order to validate model performance. The ML model was able to precisely identify the two touching OBSCs in this validation image. (B) Three examples of the ML model creating OBSC masks for experiments analyzed in this paper. (C) A step-by-step representation of the D4T measurement process. OBSC masks are generated from the brightfield component of the D4T image, and these masks are bisected to distinguish between two hemispheres of the OBSCs. Thresholding is performed within each hemisphere to identify and measure the tumor present. (D) An example of how the process shown in (C) can identify a wide range of tumor sizes from the same image. The 3D surface plot highlights the 100-fold difference in signal magnitude between Microtumors 3 (left) and 20 (right). (E) OBSC masks generated by the same ML model can be used to measure Propidium Iodide fluorescence, a marker of cell death, across entire OBSCs. (F) Triangle thresholding on D1T images precisely distinguishes signal from background. The enhanced images feature adjustments to brightness and contrast to demonstrate how thresholding can detect features that might be obscured by computer display settings. Visualization of computer vision applied to various tumor signals. (A) Two of the labeled OBSCs withheld from the ML training process in order to validate model performance. The ML model was able to precisely identify the two touching OBSCs in this validation image. (B) Three examples of the ML model creating OBSC masks for experiments analyzed in this paper. (C) A step-by-step representation of the D4T measurement process. OBSC masks are generated from the brightfield component of the D4T image, and these masks are bisected to distinguish between two hemispheres of the OBSCs. Thresholding is performed within each hemisphere to identify and measure the tumor present. (D) An example of how the process shown in (C) can identify a wide range of tumor sizes from the same image. The 3D surface plot highlights the 100-fold difference in signal magnitude between Microtumors 3 (left) and 20 (right). (E) OBSC masks generated by the same ML model can be used to measure Propidium Iodide fluorescence, a marker of cell death, across entire OBSCs. (F) Triangle thresholding on D1T images precisely distinguishes signal from background. The enhanced images feature adjustments to brightness and contrast to demonstrate how thresholding can detect features that might be obscured by computer display settings. Comparing automated image analysis to previously published manual data We next applied this image analysis workflow to a set of experiments from a previously published paper15. This published dataset included tumor-killing experiments performed on microtumors (from both established cell lines and patient tumor samples) co-cultured with OBSCs, as well as the corresponding OBSC toxicity assays. Manual measurements, which include labeling regions of interest on quantitative image files by hand, were not treated as a gold standard because they are inherently inconsistent due to human subjectivity. This subjectivity arises from variability in how different users define regions of interest (ROIs), especially in images with low signal-to-noise ratios or ambiguous boundaries. Even small differences in ROI placement can lead to substantial changes in calculated survival, particularly when signal intensity is close to background levels. Instead, we validated automation by assessing reproducibility and consistency across replicate experiments. We compared differences between the published, manually calculated survival of each microtumor and OBSC and our new automated measurements by generating Bland-Altman graphs (Figs. 3A and 4D). Bland-Altman graphs were chosen for this analysis because, compared to simpler correlation calculations and plots, they provide a more accurate assessment of how much two measurement methods disagree and whether the degree of disagreement is acceptable19. For each microtumor and OBSC, the average of the manual and automated measurements is plotted on the x-axis. The difference between the two measurements (Automated - Manual) is plotted on the y-axis. The mean difference is plotted as a black line, and the 95% confidence interval, conveying the distribution of differences, is plotted as dotted gray lines. The distinction between two measurements “agreeing” and “disagreeing” is inherently arbitrary; in this study, we defined a difference of 10% survival as our agreement threshold. Measurements differing by less than 10% were considered to be in agreement, while those differing by more than 10% were further investigated to determine whether automation had improved accuracy or precision compared to the manual results. Across 886 microtumors, the manual and automated survival measurements showed a mean difference of −1.69% and a 95% Confidence Interval of (−30.9%, 27.59%) (Fig. 3A). 723 of these measurements (82%) were within our 10% agreement threshold, while 163 measurements (18%) differed by more than 10%. In the experiment DIPG vs. Etoposide 3/31/2022 (Fig. 3A, pink), automation produced several differing survival measurements, including one that was 183% lower than the corresponding manual measurement. A dose–response plot comparing Manual and Automated measurements for this experiment revealed that the automated values for microtumors 11 and 15 are more consistent with the survival of other replicate microtumors at those doses (Fig. 3B). Automation reduced the standard deviation from 90.1% to 41.9% in the 50 µM dose group and from 116.2% to 42.5% in the 100 µM dose group. Inspection of the automatically generated masks for these microtumors confirms that automation captured the signals precisely (Fig. 3C). These discrepancies likely stem from the inherent subjectivity of manual ROI placement, particularly in images with low signal-to-noise ratios. In such cases, even slight overestimation of ROI boundaries, such as drawing an ROI that is too large, can disproportionately inflate survival measurements, especially when bioluminescence signals are close to background levels. This sensitivity to ROI definition underscores the need for consistent, algorithmic segmentation. Four experiments, including this DIPG vs. Etoposide experiment in question, accounted for all disagreements greater than 50%, and in each of these cases, the disagreement reflected a significant improvement in accuracy through automation (Supplemental Figs. 3A-D). Comparisons of Manual and Automated measurements for all other microtumor experiments are provided in Supplemental Fig. 4. The manual and automated measurements of off-target OBSC toxicity showed an even higher level of agreement, with a mean difference of 0.75% and a 95% confidence interval of (−4.86%, 6.35%) (Fig. 3D). Here, 1,684 of 1,723 OBSCs (98%) had measurements that agreed within 10%; only 39 OBSCs (2%), all from two experiments, featured automated measurements that differed substantially from manual measurements. The experiment that showed the highest disagreement, featuring automated measurements of survival up to 26% higher than manual, was OBSC vs. Etoposide 04/04/2022 (Fig. 3D, pink). A second OBSC vs. Etoposide experiment was conducted on 05/16/2022, allowing us to evaluate automation’s performance by assessing its impact on inter-experiment consistency (Fig. 3E). The manual measurements for these two experiments (Fig. 3E, blue) differed substantially at every shared dose group (50 µM: p = 0.051; 250 µM: p = 0.001; 1000 µM: p = 0.027), whereas automation led to remarkable consistency between the experiments (50 µM: p = 0.300; 250 µM: p = 0.851; 1000 µM: p = 0.751) (Fig. 3E, pink). The shift across all dose groups can be attributed to shifts in the control group measurements. Since each dose group is normalized to the control group, it is crucial to obtain accurate control measurements. The automated masks for the negative control group in the 04/04/2022 experiment aligned with the fluorescent signals upon visual inspection (Fig. 3F), providing additional confidence that the shift in results produced by automation reflects improved accuracy. While manual measurements were not necessarily erroneous, they are inherently variable due to subjective ROI placement, especially when signal intensity is near background. Automation improves reproducibility and provides a transparent record of how measurements were generated. The only other OBSC experiment with disagreements greater than 10% between manual and automated measurements also featured improvements in the control group that affected all dose groups (Supplemental Figs. 3E-F). Fig. 3Comparing automated image analysis measurements to previous manual measurements. (A) A Bland-Altman graph comparing Manual and Automated Image Analysis results for each microtumor. The difference between Automated – Manual measurement is plotted on the y-axis, and the average between the two is plotted on the x-axis. The experiment highlighted in pink is DIPG vs. Etoposide 03/31/2022. (B) A Dose–response plot for DIPG vs. Etoposide 03/31/2022 comparing the manual and automated results. Microtumors where Auto IA produced a large difference are highlighted. (C) The automatically generated masks for the highlighted microtumors above (11 on left, 15 on right). Microtumor 15 displays a low SNR, making an accurate ROI crucial. (D) A Bland-Altman graph comparing Manual and Automated Image Analysis results for each OBSC in the OBSC toxicity assays. The experiment highlighted in pink is OBSC vs. Etoposide 04/04/2022. OBSCs from positive control groups lie near 0% survival. (E) A Dose–response graph showing how automation (pink) improves inter-experiment consistency between replicate experiments conducted in April (circles, solid line) and May (triangles, dashed line). (F) The automatically generated OBSC masks for the selected OBSC toxicity assay. Comparing automated image analysis measurements to previous manual measurements. (A) A Bland-Altman graph comparing Manual and Automated Image Analysis results for each microtumor. The difference between Automated – Manual measurement is plotted on the y-axis, and the average between the two is plotted on the x-axis. The experiment highlighted in pink is DIPG vs. Etoposide 03/31/2022. (B) A Dose–response plot for DIPG vs. Etoposide 03/31/2022 comparing the manual and automated results. Microtumors where Auto IA produced a large difference are highlighted. (C) The automatically generated masks for the highlighted microtumors above (11 on left, 15 on right). Microtumor 15 displays a low SNR, making an accurate ROI crucial. (D) A Bland-Altman graph comparing Manual and Automated Image Analysis results for each OBSC in the OBSC toxicity assays. The experiment highlighted in pink is OBSC vs. Etoposide 04/04/2022. OBSCs from positive control groups lie near 0% survival. (E) A Dose–response graph showing how automation (pink) improves inter-experiment consistency between replicate experiments conducted in April (circles, solid line) and May (triangles, dashed line). (F) The automatically generated OBSC masks for the selected OBSC toxicity assay. Multistep approach to Dose–response modeling allows for robust automation After microtumor kill and off-target OBSC toxicity are measured, we estimate clinically relevant features of dose–response behavior (e.g., Effective Doses) for downstream input into the overall DSS. Previously, we used linear interpolation between dose groups to estimate such parameters. However, linear interpolation only considers two dose groups at a time, leaving it vulnerable to local variations in data20. Dose–response modeling (DRM), which fits a parametric model to an entire data set, provides a more biologically plausible and robust understanding of treatment efficacy20,21. The open-source dose–response modeling package drc, written in R, features a wide variety of models and functions that enable rapid and accurate model fitting22. When using DRM instead of linear interpolation, the challenge is selecting an appropriate model for the data20. Researchers often inspect a dataset and either select a model they expect to fit or compare the fits of several candidate models. We automated this process by developing a multistep program that emulates an intuitive approach to DRM (Fig. 4). Heuristic categorization of dose response behavior Our microtumor and OBSC survival data exhibited a variety of dose–response behaviors, some more complex than others. To ensure appropriate model fitting across all datasets, we heuristically categorized experiments into three categories, applying a different set of models for each category (Fig. 4A-C; Supplemental Figs. 5–6). In “Decreasing” experiments, higher doses led to increased tumor or OBSC kill until a final asymptote was reached (Fig. 4A). Three families of simple models were tested on decreasing experiments. An experiment was placed in the “Plateau” category if a plateau separated two regions of killing (Fig. 4B), where a plateau was defined as three or more dose groups in which survival changed by less than 15% between doses. In these experiments, the Cedergreen-Ritz-Streibig (CRS) model was considered in addition to the Monotonic Decreasing models23. If the plateau effect was minor, a simpler decreasing model could still fit the data, thereby avoiding overfitting. “Growth” data exhibited more than 10% tumor growth at one or more doses (Fig. 4C), and for these experiments, we added the Brain-Cousens (BC) model to the CRS model and the list of decreasing models24. Dynamically defining model parameters Consistent automated model fitting required careful fine-tuning of parameter estimation. We defined a range of possible values for each parameter to ensure expected behavior and we fixed certain parameters to specific values based on our experimental constraints (e.g., survival at Dose = 0 was always 100%). We also defined certain parameter values dynamically based on each experiment’s data. For example, the default parameterization of the BC and CRS models provides limited control over hormesis behavior (initial growth at low doses followed by killing at high doses), since no single parameter governs the magnitude of the hormesis effect. Occasionally, these models failed to accurately capture the hormesis effect, particularly in cases where only one or two doses exhibited growth (Fig. 4D). To improve modeling of hormesis, we implemented a model that uses a Reparametrized version of the BC model (RBC)25. This model included a parameter representing the dose at which maximum growth occurred. Defining this parameter dynamically based on each experiment’s survival data enabled the RBC model to faithfully capture hormetic behavior while still fitting the subsequent tumor kill (Fig. 4E). A dynamic approach was also used to fit exponential decay models consistently. The model fitting process requires initial parameter estimates, and the quality of these guesses can determine whether a model fits successfully. By using a custom method based on linear interpolation (Supplemental Fig. 7), we generated initial parameter estimates that resulted in consistent model fitting. Selecting the final model After selecting appropriate candidate models, we used a modified version of drc’s function mselect to fit and compare them. Mselect calculates two statistics, which we combined to select the final model. Lack of Fit measures how closely the given model aligns with the data points, and a value close to 1 is ideal26. Akaike’s Information Criterion (AIC) considers model complexity, favoring simpler models with fewer parameters27. Lower AIC values indicate more suitable models. In our dataset, we identified cases where AIC values were similar between two models, but Lack-of-Fit indicated that one model aligned more closely with the underlying data. To account for this, we combined the two criteria (see Methods), allowing the more discriminative measure to dominate (Fig. 4F). This approach selected models that aligned with the underlying data while still accounting for model complexity (Fig. 4G). Fig. 4Automated selection of dose–response models. Through a combination of heuristics, rational parameter choices, and plausibility checks, the model fitting process could be controlled and customized for each data set. (A) Monotonic Decreasing data fit by a Weibull model (pink). (B) Biphasic data fit by a CRS model (pink). None of the standard decreasing models could accommodate all of the “plateau” points, and none reached the complete response seen at the highest dose. (C) Growth data displaying hormesis behavior fit by the standard Brain-Cousens model (pink). (D) Survival Data displaying hormesis behavior where standard models did not fit well to the data. To capture the hormesis behavior accurately, we found the dose of maximum growth and used it while fitting the Reparametrized Brain-Cousens model (RBC). (E) In this example, dynamically defining a parameter while fitting RBC led to a model that captured hormesis behavior without “overshooting” the observed peak survival. (F) This plot visualizes how we combined AIC and Lack-of-Fit to select the best fit model. A darker blue corresponds to a better Combined Criterion score. Each point corresponds to the scores of a given fitted model. In this example (U373KO vs. Trametinib), the chosen model, CRS.4c, had the best Combined score, even though it had a slightly lower AIC than the RBC model. (G) Qualitative confirmation that the selected CRS.4c model fits the underlying survival data better than RBC.5. Automated selection of dose–response models. Through a combination of heuristics, rational parameter choices, and plausibility checks, the model fitting process could be controlled and customized for each data set. (A) Monotonic Decreasing data fit by a Weibull model (pink). (B) Biphasic data fit by a CRS model (pink). None of the standard decreasing models could accommodate all of the “plateau” points, and none reached the complete response seen at the highest dose. (C) Growth data displaying hormesis behavior fit by the standard Brain-Cousens model (pink). (D) Survival Data displaying hormesis behavior where standard models did not fit well to the data. To capture the hormesis behavior accurately, we found the dose of maximum growth and used it while fitting the Reparametrized Brain-Cousens model (RBC). (E) In this example, dynamically defining a parameter while fitting RBC led to a model that captured hormesis behavior without “overshooting” the observed peak survival. (F) This plot visualizes how we combined AIC and Lack-of-Fit to select the best fit model. A darker blue corresponds to a better Combined Criterion score. Each point corresponds to the scores of a given fitted model. In this example (U373KO vs. Trametinib), the chosen model, CRS.4c, had the best Combined score, even though it had a slightly lower AIC than the RBC model. (G) Qualitative confirmation that the selected CRS.4c model fits the underlying survival data better than RBC.5. Overall impact of automation on drug sensitivity scores Automatically fitted models were used to calculate dose–response parameters, which were then combined into a single DSS for each therapeutic against each tumor. Automation also facilitated consistent, statistically driven methods for calculating dose–response parameters from the models and the underlying survival data. To determine the impact of automation on DSS calculation, we calculated automated DSSs for 31 previously conducted tumor kill experiments, 20 using tumor lines and 11 using patient tumor tissues. We first evaluated the total change in DSS from the published manual process to the new automated workflow and then isolated the changes resulting from automating image analysis and dose–response modeling, respectively. In 26 out of 31 experiments, automation produced a change in DSS of less than 18 points (< 10% of the total DSS range, which exists from − 100 to + 100), with no clear bias toward increasing or decreasing DSS scores (Mean Change = −0.03 points) (Fig. 5, top row). Automating image analysis (Fig. 5, middle row) only caused two experiments to shift by more than 18 points, while automating DRM and DSS calculation (Fig. 5, bottom row) shifted three experiments by more than 18 points. Six experiments were selected to illustrate how automation impacts DSS calculation in greater detail (Fig. 6; larger images available in Supplemental Figs. 8–13). In each panel, the previously published manual results are shown in the left column, and the new automated results are shown in the right. The top graphs display dose–response curves fitted to the survival data, while the bottom waterfall plots visualize how the DSS is calculated from its component parameters. For DIPG vs. Etoposide (Fig. 5, pink), which was featured earlier in this paper (Fig. 4A-C), automated image analysis measured less tumor growth at low doses, increasing the DSS by 41 points (Fig. 6A). In the second highlighted experiment, LN229 vs. Etoposide, the original manual technician qualitatively determined that LN229 survival did not change substantially across the last three doses, so the tumor response was classified biphasic (Fig. 6B). However, the automated DSS process evaluates whether a statistically significant difference exists among the last three dose groups. The 1000µM group was statistically significantly different from the previous two dose groups, so the response was classified as not biphasic; this drove the 13-point increase in DSS. The next two highlighted experiments exemplify how automated DSSs closely matched the earlier manual results in many cases (Fig. 6C, D). The experiment testing PT220029 vs. Radiation demonstrates how automated DRM and DSS calculation can affect the final DSS even when automated and manual image analyses matched (Fig. 6E). Here, the automated models yielded a smaller AUC window. The models also showed that, while some fluctuation exists in the underlying survival data, both the Tumor and OBSC responses plateaued at the highest doses, leading to no points being added for a difference in Slope. The final highlighted experiment evaluated Trametinib’s toxicity against PT210508, a patient-derived sample of Grade 2 Multiple Meningioma (Fig. 5, white). Here, the overall shift in DSS from 48.4 points (Manual) to 19.7 points (Automated) was driven by changes in how the Slope and Biphasic parameters were calculated (Fig. 6F). The automated models also produced improved estimates of Effective Doses, which further decreased the DSS. Fig. 5Automation’s impact on DSS scores. A bar graph demonstrating how automation changed DSS results across 31 OBSC platform experiments. All rows feature experiments in the same order, and all values represent the change from Manual to Automated. Six highlighted experiments will be explored further in Fig. 6. Top Row: The change in DSS observed when Image Analysis is automated. Middle Row: The change in DSS observed when Automatic Dose–response Models are fit to data and used to calculate DSS. Bottom Row: The total change in DSS from the completely manual calculations to the completely automated workflow. Automation’s impact on DSS scores. A bar graph demonstrating how automation changed DSS results across 31 OBSC platform experiments. All rows feature experiments in the same order, and all values represent the change from Manual to Automated. Six highlighted experiments will be explored further in Fig. 6. Top Row: The change in DSS observed when Image Analysis is automated. Middle Row: The change in DSS observed when Automatic Dose–response Models are fit to data and used to calculate DSS. Bottom Row: The total change in DSS from the completely manual calculations to the completely automated workflow. Fig. 6automation’s impact on individual experiments. An investigation of how automation impacted the DSS results in six experiments. Each experiment features four panels. Top Left: manual image analysis and linear interpolation. Bottom Left: a waterfall plot showing the manual DSS parameters (pink, light blue) and how they add up to the final DSS (navy). Top Right: Automated Image Analysis and Dose–response Models. Bottom Right: A waterfall plot for the automated parameters and DSS. (A) DIPG vs. Etoposide, also featured in Fig. 4. (B) LN229 vs. Etoposide (C) MS21 vs. Cisplatin (D) PT210468 vs. TR107 (E) PT220029 vs. Radiation (F) PT210508 vs. Trametinib. automation’s impact on individual experiments. An investigation of how automation impacted the DSS results in six experiments. Each experiment features four panels. Top Left: manual image analysis and linear interpolation. Bottom Left: a waterfall plot showing the manual DSS parameters (pink, light blue) and how they add up to the final DSS (navy). Top Right: Automated Image Analysis and Dose–response Models. Bottom Right: A waterfall plot for the automated parameters and DSS. (A) DIPG vs. Etoposide, also featured in Fig. 4. (B) LN229 vs. Etoposide (C) MS21 vs. Cisplatin (D) PT210468 vs. TR107 (E) PT220029 vs. Radiation (F) PT210508 vs. Trametinib. Discussion Automation of image and data analysis significantly increased the translational potential of the OBSC functional diagnostic platform by improving accuracy, consistency, objectivity, and efficiency. Implementation of computer vision either maintained or improved the accuracy of measurements, as demonstrated by global observations and individual case studies. A multistep approach to Dose–response Modeling, balancing simplicity and detail, successfully fit models across a variety of response behaviors. The resulting Drug Sensitivity Scores (DSSs), derived from these models and statistical tests on the raw data, more accurately represented the relationship between treatment and tumor. Automation reduced analysis time from approximately 20 h per ten-drug drug sensitivity assay to just 15 min. Automation also mitigates previously observed variability in manual, subjective measurements by applying consistent, objective algorithms across all datasets. Our functional therapeutic efficacy data are stored in an organized database that facilitates future investigations into the clinical relevance of the OBSC platform. The primary impact of this work is efficiency: automation reduces analysis time by 99%, enabling researchers to process more patient samples and expand clinical feasibility studies. Beyond efficiency, improved accuracy and reproducibility enhance the reliability of functional precision medicine data, which is critical for correlating ex vivo drug responses with patient outcomes. By generating more consistent data across a larger number of patient tumors, this workflow enhances our ability to identify clinically meaningful patterns and predictive biomarkers in future studies. Some analysis automation programs, such as the computer vision software CellProfiler5, offer general modules that can be used for a wide variety of applications. Biodock fits in this category as well, making Machine Learning accessible for various biological imaging applications. Other programs fill more specific niches that are directly useful to broader communities of researchers. MyoSight3 analyzes cross-sections of muscle tissue; Neurite-J7 and Neurite Tracer8 both measure axonal growth in neurites. However, many assays, including our own, require custom solutions that account for the intricacies of the novel technologies and applications. While our workflow is customized to our assay, the approaches used in this paper are generalizable to other applications. The integration of Biodock via API combined the capabilities of machine learning with the flexibility of executing custom scripts in a local Jupyter Notebook, making ML accessible to a broader range of scientists. Training our ML model on brightfield structures allowed us to apply our model in two separate workflows and to account for large dynamic ranges and “missing” signals; benefits that could be realized by any researcher using multimodal imaging systems such as AMI or IVIS. While pixel-level accuracy metrics for tumor segmentation were not feasible due to the absence of ground truth masks, our validation strategy focused on reproducibility and biological plausibility. Automated measurements reduced variance across replicates and improved inter-experiment consistency, outcomes that are critical for translational applications. For OBSC segmentation, quantitative validation on a separate dataset confirmed high accuracy, and visual inspection of experimental images supported the generalizability of these results. This approach of comparing automated and manual results and systematically investigating large differences could be adapted to validate other automation workflows in the absence gold-standard measurements. Beyond improving consistency and efficiency, automation revealed biologically relevant features that were previously obscured by manual subjectivity. For example, in DIPG vs. Etoposide (Fig. 6A), automated analysis detected less apparent tumor growth at low doses, which increased the DSS by 41 points. This suggests that manual interpretation resulted in ROIs that were slightly too large but had large downstream, potentially misrepresenting tumor resilience at subtherapeutic concentrations. Accurate characterization of early tumor behavior (e.g., EC10 and EC25) is critical as these low-dose responses may correlate with clinical sensitivity and therapeutic window. Similarly, in LN229 vs. Etoposide (Fig. 6B), automation reclassified the response from “biphasic” to “not biphasic” based on statistical criteria. This correction eliminated an artificial plateau introduced by subjective judgment and revealed a more biologically plausible monotonic kill curve. More broadly, the ability to model hormesis accurately, enabled by dynamic parameterization of the Reparametrized Brain-Cousens model, provides new insight into tumor-drug interactions. Hormetic behavior, where low doses stimulate growth before higher doses induce killing, may indicate adaptive stress responses or compensatory signaling pathways. Identifying these patterns consistently could inform combination strategies that suppress hormesis and improve therapeutic efficacy. While automated model fitting improves objectivity and reproducibility, it does not replace biological interpretation and robust experimental design. Our workflow assumes that bioluminescent signal intensity is proportional to the number of viable tumor cells, regardless of morphological changes, and our implementation of control groups depends on this assumption. This assumption may not capture phenomena such as edema or diffuse invasion, which could mimic growth or shrinkage without reflecting true therapeutic response. Similarly, hormetic behavior detected at low doses is interpreted as a drug-induced effect, but alternative explanations (e.g., osmotic responses) cannot be excluded without additional biological validation. These limitations underscore that automated analysis should complement, not replace, expert review and integration with biological context. Future developments will continue to streamline the workflow and improve results. Our workflow currently lacks a comprehensive User Interface (UI), but we are developing one to further enhance the efficiency and accessibility of automation. This app will also feature a more modular codebase that can be easily adapted to new OBSC-based workflows, such as immune cell therapy or longitudinal growth studies. Automating DSS calculation revealed that parameters with discrete categories (e.g., growth vs. no growth) sometimes led to unexpected DSS differences between similar-looking data. Making these parameters continuous would lead to more consistent and linear DSS calculations. When designing our model-fitting approach, we aimed to faithfully represent tumor behavior while minimizing the risk of overfitting; however, we have not empirically confirmed that our approach achieves an ideal balance between the two. Measuring how a tumor responds to previously untested doses would reveal whether our approach accurately interpolates between dose groups. An exciting future application of this workflow is the integration of brain region mapping to investigate how tumor growth and therapeutic response vary by anatomical context. Tumor–neuron interactions are known to influence invasion and drug sensitivity, and OBSC slices retain regional architecture that could capture these effects. Extending our computer vision approach to classify slices by brain region, using features such as size, shape, and texture, would enable correlation of drug response with neuroanatomical context. This capability could reveal region-specific vulnerabilities and inform precision therapy strategies. Finally, the relative clinical importance of each DSS parameter remains unclear. Translation of the DSS will require investigating the predictive power of various dose–response parameters in a clinical context. Our automated analysis workflow quickly generates accurate and consistent results. The time saved by automation frees researchers to process higher volumes of patient tissue, interpret results, and plan new experiments. It positions the OBSC platform for in-depth investigation into genetic and clinical correlations, maximizing its potential to uncover the utility of OBSC–tumor co-cultures as a clinically relevant model. Automation has thus advanced the OBSC assay toward our lab’s broader goal, and that of the fieldof meaningfully improving patient outcomes. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1