AMA-SAM: Adversarial Multi-Domain Alignment of Segment Anything Model for High-Fidelity Histology Nuclei Segmentation

Jiahe Qian, Yaoyu Fang, Jinkui Hao, Bo Zhou · 2025

Read AMA-SAM: Adversarial Multi-Domain Alignment of Segment Anything Model for High-Fidelity Histology Nuclei Segmentation by Jiahe Qian, Yaoyu Fang, Jinkui Hao, Bo Zhou (2025) 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-2503-21695.txt.

Introduction

Jiahe Qian, Yaoyu Fang, Jinkui Hao, and Bo Zhou present AMA-SAM, a 2025 method for outlining cell nuclei in tissue images. Clear outlines help doctors and researchers read such scans. Most tools train on one data set, yet scans from other labs differ, and mixing them can hurt results more than it helps.

Their method adapts a general image model called SAM in two ways. One part aligns traits from many data sets so the model learns what stays the same across labs. A second part draws maps at full sharpness, so fine edges of nuclei stay clear. On several public data sets, the method beat prior tools by a clear margin. It shows how one model can learn from many scan sources at once.

Catalog record

CREATORJiahe Qian, Yaoyu Fang, Jinkui Hao, Bo Zhou
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-21695

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

AMA-SAM: Adversarial Multi-Domain Alignment of Segment Anything Model for High-Fidelity Histology Nuclei Segmentation Jiahe Qian, Yaoyu Fang, Jinkui Hao, Bo Zhou Abstract Accurate segmentation of cell nuclei in histopathology images is essential for numerous biomedical research and clinical applications. However, existing cell nucleus segmentation methods only consider a single dataset (i.e., primary domain), while neglecting to leverage supplementary data from diverse sources (i.e., auxiliary domains) to reduce overfitting and enhance the performance. Although incorporating multiple datasets could alleviate overfitting, it often exacerbates performance drops caused by domain shifts. In this work, we introduce Adversarial Multi-domain Alignment of Segment Anything Model (AMA-SAM) that extends the Segment Anything Model (SAM) to overcome these obstacles through two key innovations. First, we propose a Conditional Gradient Reversal Layer (CGRL), a multi-domain alignment module that harmonizes features from diverse domains to promote domain-invariant representation learning while preserving crucial discriminative features for the primary dataset. Second, we address SAM’s inherent low-resolution output by designing a High-Resolution Decoder (HR-Decoder), which directly produces fine-grained segmentation maps in order to capture intricate nuclei boundaries in high-resolution histology images. To the best of our knowledge, this is the first attempt to adapt SAM for multi-dataset learning with application to histology nuclei segmentation. We validate our method on several publicly available datasets, demonstrating consistent and significant improvements over state-of-the-art approaches. 1 Introduction Nuclei segmentation is a critical task in digital pathology analysis, playing a central role in disease diagnosis, histological research, and personalized medicine. In tissue sections and microscopic images, cell nuclei’s morphology, size, and spatial distribution provide vital information for assessing tissue health, identifying pathological alterations, and predicting disease progression. Although deep learning and computer vision have enabled the development of numerous neural network-based approaches for histology nuclei segmentation [ [ 32 , 50 , 49 , 13 , 33 , 23 , 3 , 12 , 14 , 5 , 31 , 15 ] ], existing methods are trained on a single dataset, rendering them prone to overfitting to the specific characteristics of that dataset. Moreover, a naive fusion of multiple datasets often leads to further performance degradation because differences in imaging acquisition systems, staining protocols, and tissue types across datasets introduce substantial domain shifts, which can confound the learning process when data are simply mixed. Consequently, developing a robust histology nuclei segmentation model that can effectively leverage multi-source data to enhance performance on the target dataset (i.e. primary dataset/domain), while efficiently leveraging supplementary data from diverse sources (i.e. auxiliary domains) remains a significant challenge, with no universal solution established to date. The Segment Anything Model (SAM) [ [ 22 ] ] represents a significant advancement in universal image segmentation, owing to its robust architecture and large-scale pretraining that enable remarkable adaptability and accuracy across a broad spectrum of natural images. Despite these advances, SAM exhibits some limitations in histological nuclei segmentation. First, SAM employs a fixed output resolution of 256\times 256 pixels, necessitating upsampling of segmentation outputs via interpolation to match the original image’s higher resolution required for detailed histology analysis. This process can blur nuclei boundaries and adversely affect downstream analytical tasks, such as the accurate computation of area or perimeter especially for small nuclei. Secondly, while SAM can be fine-tuned on diverse datasets, the persistent challenge of domain discrepancy arising from differences in image characteristics and annotation standards continues to compromise its performance when training on multiple datasets. These limitations underscore the need for enhanced methodologies that can effectively handle high-resolution inputs and reconcile domain differences to achieve robust segmentation outcomes. To address these challenges, we propose an innovative SAM-based framework, called Adversarial Multi-Domain Alignment of Segment Anything Model (AMA-SAM), which leverages collaborative training on a primary dataset and auxiliary datasets from diverse sources to achieve high-fidelity high-resolution histological nuclei segmentation on the primary dataset. Specifically, our contributions include: (1) Multi-domain Alignment via Conditional Gradient Reversal Layer (CGRL) : We introduce a novel CGRL for SAM to facilitate effective adversarial multi-domain alignment. By selectively applying gradient reversal, the CGRL aligns features from auxiliary domains to the primary domain while preserving the integrity of the primary domain’s features. This targeted alignment enables the integration of complementary information from auxiliary datasets, thereby significantly enhancing segmentation performance on the primary dataset. (2) High-Resolution Decoder (HR-Decoder) : To address SAM’s low output resolution, we design an HR-Decoder that generates high-resolution segmentation results. By freezing SAM’s original decoder and incorporating additional slice tokens, along with our Multi-token Slices Producer and Pixel Ensemble Module, our method reconstructs detailed high-resolution segmentation, preserving fine segmentation boundaries and structural details. (3) Comprehensive Validation : We rigorously evaluate AMA-SAM across a variety of datasets and experimental scenarios. The results demonstrate the framework’s superior segmentation performance on the primary dataset, achieved by effectively integrating complementary information from auxiliary datasets. Our proposed framework combines innovative domain alignment strategies with architectural enhancements to overcome key limitations of SAM in high-resolution histological nuclei segmentation. 2 Related Work Nuclei Segmentation: Early methods for nuclei segmentation predominantly relied on handcrafted features and classical image processing techniques [ [ 45 , 1 , 27 ] ]. With the advent of deep learning, a multitude of convolutional neural network-based approaches have been developed for histology nuclei segmentation [ [ 32 , 50 , 49 , 13 , 33 , 23 , 3 , 12 , 14 , 5 , 31 , 15 , 36 , 4 , 40 , 18 , 8 ] ]. For instance, HoVer-Net [ [ 13 ] ] addresses the problem of overlapping nuclei by predicting horizontal and vertical displacement maps, whereas Micro-Net [ [ 33 ] ] enhances the classical U-Net architecture [ [ 34 ] ] with multi-scale feature extraction to capture fine morphological details. Semi-automatic methods such as NuClick [ [ 23 ] ] incorporate user-provided scribbles for guidance, and more recent approaches including CDNet [ [ 14 ] ], CPP-Net [ [ 5 ] ], and PROnet [ [ 31 ] ] further improve boundary delineation by employing directional feature maps, complementary boundary masks, or offset maps. More recently, UN-SAM [ [ 8 ] ] fine-tuned SAM using an automatic learning prompt strategy, achieving state-of-the-art performance in nuclei segmentation on single dataset. Multi-dataset training: Multi-dataset training [ [ 37 , 44 , 46 , 28 , 29 ] ] has emerged as a promising strategy to enhance model generalizability and robustness by leveraging data from heterogeneous sources. Prior studies have demonstrated that integrating diverse datasets can mitigate dataset bias [ [ 17 ] ] and boost performance in tasks such as object detection [ [ 48 , 41 , 42 , 7 , 47 , 38 , 6 , 43 ] ] and semantic segmentation [ [ 2 , 11 , 26 ] ]. However, existing methods of multi-dataset co-training exhibit specific limitations that render them suboptimal for histology nuclei segmentation tasks. First, most existing methods for multi-dataset co-training in segmentation primarily address label inconsistencies across datasets [ [ 2 , 35 , 25 ] ], rather than resolving domain gaps of images. Second, existing multi-dataset training approaches typically treat all datasets equally to achieve balanced improvements. In practice, significant variability in data characteristics leads to a dilution of performance on the primary dataset compared to

Read the complete text