Biologically-inspired Semi-supervised Semantic Segmentation for Biomedical Imaging
Luca Ciampi, Gabriele Lagani, Giuseppe Amato, Fabrizio Falchi · 2024
Read Biologically-inspired Semi-supervised Semantic Segmentation for Biomedical Imaging by Luca Ciampi, Gabriele Lagani, Giuseppe Amato, Fabrizio Falchi (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-03192.txt.
Introduction
Ciampi, Lagani, Amato, and Falchi propose a bio inspired method for medical image segmentation in twenty twenty-four. Segmentation labels each pixel by tissue or structure type. Training such models needs many labelled images, which are scarce in medicine. Their method learns first without labels, then fine tunes on a small labelled set.
The first stage uses a Hebbian rule, summed up as fire together, wire together. It updates both downsampling and upsampling layers without backpropagation. The second stage fine tunes the model with standard methods on few labels. Tests on several biomedical data sets beat state of the art methods at varied label levels. Using their first stage to start rival methods also improved those methods. Code to repeat the tests is shared.
Catalog record
| CREATOR | Luca Ciampi, Gabriele Lagani, Giuseppe Amato, Fabrizio Falchi |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-03192
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Biologically-inspired Semi-supervised Semantic Segmentation for Biomedical Imaging Luca Ciampi, Gabriele Lagani, Giuseppe Amato, Fabrizio Falchi Abstract We propose a novel bio-inspired semi-supervised learning approach for training downsampling-upsampling semantic segmentation architectures. The first stage does not use backpropagation. Rather, it exploits the Hebbian principle “fire together, wire together” as a local learning rule for updating the weights of both convolutional and transpose-convolutional layers, allowing unsupervised discovery of data features. In the second stage, the model is fine-tuned with standard backpropagation on a small subset of labeled data. We evaluate our methodology through experiments conducted on several widely used biomedical datasets, deeming that this domain is paramount in computer vision and is notably impacted by data scarcity. Results show that our proposed method outperforms SOTA approaches across different levels of label availability. Furthermore, we show that using our unsupervised stage to initialize the SOTA approaches leads to performance improvements. The code to replicate our experiments can be found at https://github.com/ciampluca/hebbian-bootstraping-semi-supervised-medical-imaging Journal: Neural Networks 1 Introduction Semantic segmentation in biomedical images assigns pixel-level class labels to structures like cells, tumors, and lesions, playing a key role in automating computer-aided diagnoses. Recent data-driven deep-learning approaches using CNNs and Transformers have shown excellent results [ 1 , 2 , 3 , 4 , 5 , 6 , 7 ] . However, these methods demand extensive annotated data for supervised backpropagation-based training, which restricts their large-scale application despite their significance in computer vision [ 8 , 9 , 10 , 11 , 12 , 13 ] . Nevertheless, integration between biological mechanisms and deep learning (BIDL) seems a promising direction, tackling not only the substantial demand for data but also allowing complex neural computation with extreme energy efficiency [ 14 , 15 ] . In particular, the so-called Hebbian principle represents a simple local learning rule that closely mimics the synaptic adaptations of brain mechanisms [ 16 , 17 , 18 ] , offering an appealing and still relatively unexplored solution for extracting data features without relying on supervised backpropagation and on the availability of large labeled datasets. Different Hebbian learning strategies involving Soft-Winner-Takes-All (SWTA) [ 19 , 20 , 21 ] , or Principal Component Analysis (HPCA) [ 22 , 23 , 24 ] , allow a group of neurons to discover unsupervised features from a set of data such as clusters or principal components, and can be used to fulfill unsupervised and semi-supervised learning strategies coming to the rescue to mitigate data scarcity issues. In this work, we introduce a two-stage semi-supervised biologically-inspired learning approach for semantic segmentation in biomedical images (Fig. 1 ). In the first stage, we exploit a set of unlabeled data to train downsampling-upsampling semantic segmentation architectures in an unsupervised fashion using the Hebbian learning principle implemented with different strategies. In the second step, we grab these weights to initialize the model, and we fine-tune it through standard backpropagation supervised training on a few labeled data samples. Specifically, during the unsupervised round, we employ Hebbian learning rules in standard convolutional layers in the downsampling path and, for the first time, we formulate novel Hebbian learning strategies for transpose-convolutional (T-Conv) layers exploited for upsampling the feature maps. Figure 1 : Our semi-supervised, bio-inspired approach for semantic segmentation architectures with downsampling and upsampling paths. The method unfolds in two stages. First, we employ Hebbian learning for unsupervised pre-training on a large set of unlabeled data, enabling the model to autonomously identify features like cluster centroids. Notably, we derive new Hebbian learning rules for the transpose-convolution layers in the upsampling path. In the second stage, we apply supervised backpropagation to fine-tune the model using a small labeled dataset. We perform an experimental assessment of our methodology over three public datasets widely used in the context of medical 2D image segmentation [ 8 , 11 , 12 , 13 ] showcasing different imaging modalities and segmentation tasks, i.e., GlaS [ 25 ] for colorectal cancer segmentation in Hematoxylin and Eosin (H&E) stained histological images, PH2 [ 26 ] for skin lesion segmentation in dermoscopic images, and HMEPS [ 27 ] for eyes pupil segmentation in grayscale images. Moreover, we also conduct an analysis using the LA [ 28 ] dataset for left atrial segmentation in MRIs, thus extending our method to volumetric images. We compare our approach against several SOTA single-stage semi-supervised approaches relying on pseudo-labeling and consistency training [ 29 , 30 , 31 , 32 , 33 , 34 ] , as well as some other two-stage pipelines exploiting Variational Auto-Encoders (VAEs) [ 35 ] and Self-Supervised Learning (SSL) [ 36 ] for the unsupervised step. The outcomes demonstrate that our approach can substantially improve performance considering various label scarcity conditions. Furthermore, we show that using our unsupervised Hebbian stage to initialize the SOTA single-stage approaches results in performance improvements. Concretely, our contributions are the following: 1. We propose a novel semi-supervised two-stage pipeline for semantic segmentation, where a first unsupervised stage relying on the bio-inspired Hebbian principle is followed by a second supervised step that fine-tunes the model on a few labeled data samples. 2. We validate our methodology on several public biomedical imaging benchmarks, demonstrating that our technique can substantially improve performance compared to SOTA methods across different degrees of label availability. 3. We conduct a further experimental analysis by initializing existing SOTA semi-supervised approaches with our Hebbian unsupervised pre-training, showing performance improvements. We organize the rest of the paper as follows. In Sec. 2 , we review related work. Sec. 3 provides a brief background on Hebbian learning. In Sec. 4 , we describe our methodology, while Sec. 5 presents the experiments, discusses the results, and includes ablation studies to validate our approach. Finally, Sec. 6 concludes the paper. Additionally, A and B provide further details on the background of Hebbian learning and the implementation, respectively. 2 Related Works 2.1 Deep Learning Bio-inspired Approaches Recent Deep Learning (DL) models have achieved remarkable success across various tasks [ 37 , 38 , 39 ] . However, key challenges remain, including the substantial demand for data [ 40 ] and energy [ 41 ] . In contrast, biological intelligence appears to overcome these limitations [ 42 , 43 ] , suggesting that a tighter integration between biological mechanisms and DL (Bio-Inspired DL - BIDL) could be a promising direction [ 44 , 45 ] . A significant area within BIDL is Spiking Neural Networks (SNNs) [ 46 , 47 , 48 , 49 ] , which model neural computation in a way that more closely resembles biological neurons. In particular, neuromorphic hardware implementations based on SNNs offer complex neural computation with exceptional energy efficiency [ 50 , 14 , 15 ] . Another crucial aspect of BIDL pertains to training algorithms. Conventional DL models rely on backpropagation, which is widely considered biologically implausible [ 51 ] . Consequently, alternative learning strategies inspired by biological synaptic plasticity have gained traction. Specifically, Hebbian learning [ 16 , 17 , 52 , 24 , 21 ] provides a biologically plausible alternative for learning in DL systems [ 23 , 19 , 20 , 52 , 21 , 53 , 54 ] . 2.2 Semi-supervised Semantic Segmentation The widespread adoption of deep learning has led