XtraLight-MedMamba for Classification of Neoplastic Tubular Adenomas

TL;DR

This work tackles CRC risk stratification by improving risk assessment of low-grade tubular adenomas using an ultra-lightweight state-space model. The proposed XtraLight-MedMamba combines ConvNext shallow feature extraction, Parallel Vision Mamba back-end, SCAB fusion, and a Fixed Non-Negative Orthogonal Classifier to achieve high accuracy with only ~32k parameters. It demonstrates 97.18% accuracy and strong F1 on a curated WSIdataset of case vs. control cohorts, with Grad-CAM visualizations indicating attention to clinically meaningful nuclear and architectural features. The approach offers a scalable, interpretable, and computationally efficient tool for digital pathology-based CRC risk prediction, potentially enabling broader clinical deployment.

Abstract

Accurate risk stratification of precancerous polyps during routine colonoscopy screenings is essential for lowering the risk of developing colorectal cancer (CRC). However, assessment of low-grade dysplasia remains limited by subjective histopathologic interpretation. Advancements in digital pathology and deep learning provide new opportunities to identify subtle and fine morphologic patterns associated with malignant progression that may be imperceptible to the human eye. In this work, we propose XtraLight-MedMamba, an ultra-lightweight state-space-based deep learning framework for classifying neoplastic tubular adenomas from whole-slide images (WSIs). The architecture is a blend of ConvNext based shallow feature extractor with parallel vision mamba to efficiently model both long- and short-range dependencies and image generalization. An integration of Spatial and Channel Attention Bridge (SCAB) module enhances multiscale feature extraction, while Fixed Non-Negative Orthogonal Classifier (FNOClassifier) enables substantial parameter reduction and improved generalization. The model was evaluated on a curated dataset acquired from patients with low-grade tubular adenomas, stratified into case and control cohorts based on subsequent CRC development. XtraLight-MedMamba achieved an accuracy of 97.18% and an F1-score of 0.9767 using approximately 32,000 parameters, outperforming transformer-based and conventional Mamba architectures with significantly higher model complexity.

Figures (8)

• Figure 1: Schematic illustrating the concept of the adenoma–carcinoma sequence described by Fearon and Vogelstein genetic-alterations. Simply put, the central idea is that colorectal cancer develops in a step-wise progressive manner, starting from the precursor lesions (adenomatous polyps), eventually progressing to invasive carcinoma. Early loss of APC is associated with hyperproliferative epithelium, increasing the risk of colonic polyp formation. This is followed by an activating mutation in KRAS, which itself promotes adenomatous polyp(s) formation. Subsequent loss of tumor suppressor genes, including TP53 and DCC, contributes to tumor progression and eventual neoplastic progression towards invasive colorectal adenocarcinoma genetic-alterations. This step-wise progression from adenoma(s) to cancer typically takes over ten years to progress epidemiology.
• Figure 2: Architectural structure of XtraLight-MedMamba model for image classification task. The proposed architecture consists of ConvNeXt blocks for local morphological feature extraction with PVM layers for parallel state-space modeling, spatial and channel attention as SCAB modules and FNOClassifier to enforce fixed non-negative orthogonal decision boundaries.
• Figure 3: Unfolded ConvNext block for Shallow Feature Extraction in XtraLight-MedMamba Model.
• Figure 4: a) PVM layer in XtraLight-MedMamba for capturing of both short- and long-range spatial dependencies through parallel state space modeling b) Mamba module, a selective state space model for sequence modeling with linear computational complexity.
• Figure 5: Sample images of the dataset used: (a) H&E-stained WSI tiles from the case group consisting of tubular adenomas with low-grade dysplasia from patients who subsequently develeped CRC (b) H&E-stained WSI tiles from the control group consisting of tubular adenomas with low-grade dysplasia from patients without subsequent development of CRC.