A View-Specific Few-Shot Comparison of Capsule Networks and Standard CNN Backbones on INbreast Mammography

View at publisher Summarise with VIRAT
Authors: Khaldoon Alhusari
Cited by: 0
References: 18

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This work compares a basic Capsule Network (CapsNet) with three convolutional baselines (a compact CNN, VGG16, and ResNet50) for few-shot binary classification of INbreast mammograms. Experiments were conducted separately for craniocaudal (CC) and mediolateral-oblique (MLO) views. Few-shot setups with 10, 20, and 31 labeled samples per class were constructed and repeated across five stratified folds. The convolutional models were trained in TensorFlow/Keras (with transfer learning for VGG16 and ResNet50 using frozen backbones in the reported runs), while the CapsNet implementation was trained in PyTorch following the original routing-by-agreement formulation. Evaluation used per-fold validation and a single held-out test set, and all results are reported as mean ± standard deviation across folds. Across all views and shot regimes, the compact baseline CNN achieved the highest mean held-out test accuracy (CC: 0.702, 0.654, 0.707 for 10, 20, 31 shots; MLO: 0.571, 0.662, 0.648). CapsNet reached mid-range performance in several settings but did not surpass the compact CNN. ResNet50 (with a frozen backbone) showed lower accuracy and greater inter-fold variability. Overall, these findings suggest that in highly label-limited mammography tasks (under the specific configurations tested here) a well-regularized compact CNN or a lightweight classifier atop pretrained features remains a strong and dependable baseline. The capsule architecture merits further investigation with training strategies tuned specifically for low-data regimes. Code will be publicly released upon acceptance.

Fields of Study & Subjects

AI in cancer detection Domain Adaptation and Few-Shot Learning Digital Radiography and Breast Imaging

Publication details

Journal / source
International Conference on Image Processing and Capsule Networks (ICIPCN)
Pages
839–844
Publication date
27 January 2026
Language
EN