MSA-DCNN: A Data-Efficient Multi-Scale Deformable CNN for Medical Image Classification

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research, Computer Vision · Depth: Expert, long

Summary

MSA-DCNN, a Multi-Scale Attention Deformable Convolutional Neural Network, is proposed for data-efficient medical image classification, addressing challenges like multi-scale morphology and limited annotations. This framework integrates adaptive multi-scale sampling, within-scale saliency refinement, learned cross-scale fusion, and auxiliary self-distillation within a unified optimization scheme. Evaluated on three public benchmarks (C-NMC, PBC, ISIC-2020) and an external leukaemia hold-out set, MSA-DCNN demonstrated competitive and often superior performance against ViT and CNN baselines, and a MICCAI semi-supervised baseline, in accuracy, F1, and AUC (binary), while utilizing fewer parameters. Ablation studies confirmed the complementary contributions of its components, showing consistent gains under reduced labels and robustness under distribution shift.

Key takeaway

For Machine Learning Engineers developing medical image classification models with limited labeled data, consider MSA-DCNN. Its integrated adaptive sampling, multi-scale attention, and self-distillation improve performance and label efficiency. You can achieve competitive accuracy, F1, and AUC with fewer parameters, even under distribution shifts and class imbalance. This framework provides a strong foundation for robust, data-efficient solutions in computational healthcare.

Key insights

MSA-DCNN unifies adaptive sampling, multi-scale attention, and self-distillation for data-efficient medical image classification.

Principles

• Adaptive receptive-field optimization preserves fine-grained detail.
• Scale-consistent saliency refinement regularizes features.
• Content-aware cross-scale integration aligns representations.

Method

MSA-DCNN uses Inception stages with deformable convolutions, MCBAM for channel-spatial recalibration, and a multi-scale attention module for fusion. Self-distillation aligns shallow and deep embeddings.

In practice

• Apply deformable convolutions for heterogeneous anatomy.
• Use self-distillation to align representations with limited labels.
• Integrate multi-scale attention for content-adaptive feature fusion.

Topics

• Medical Image Classification
• Deformable Convolutional Networks
• Multi-Scale Attention
• Self-Distillation
• Data-Efficient Learning
• CNN Architectures

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Learning Task-Aware Sampling with Shared Saliency through Density-Equalizing Mappings — Dynamically redistributing computational attention based on spatial importance enhances feature extraction efficiency.
• Entropy-Guided Self-Supervised Learning for Medical Image Classification — Combining ImageNet and entropy-guided MAE pre-training via ensemble significantly boosts medical image classification accuracy and robustness.
• DRDN: Decoupled Representation Dynamic Network for From-Scratch ViT Class-Incremental Learning — DRDN improves ViT class-incremental learning by decoupling shared backbone representation preservation from task-specific discrimination.
• DB-FGA-Net: Dual Backbone Frequency Gated Attention Network for Multi-Class Brain Tumor Classification with Grad-CAM Interpretability — DB-FGA-Net offers robust, interpretable, and augmentation-free brain tumor classification using dual-backbone and frequency-gated attention.
• IEEE Transactions on Neural Networks and Learning Systems, Volume 37, Issue 3, March 2026 — The issue highlights advancements across neural networks, computer vision, and machine learning applications.
• SSMamba: A Self-Supervised Hybrid State Space Model for Pathological Image Classification — SSMamba improves pathological image classification by addressing domain shift and enhancing local-global feature learning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.