P28 - Explainability in ConvNeXt V2 for Spectral Mammography
- Event
- iCCC2026 - iCampus Cottbus Conference
2026-05-05 - 2026-05-07
Cottbus - Band
- Poster
- Chapter
- Gesundheit & Sport
- Author(s)
- W. Ghorbel, A. Siyavashi, C. Herglotz - BTU Cottbus-Senftenberg, Cottbus
- Pages
- 231 - 234
- DOI
- 10.5162/iCCC2026/P28
- ISBN
- 978-3-910600-10-2
- Price
- free
Abstract
Deep learning models can support breast cancer diagnosis in mammography, but their predictions are not explainable enough to count on their decision. Class activation mapping (CAM) techniques highlight image regions that contribute to a model’s decision, yet existing approaches frequently produce diffuse saliency maps that include large areas with limited influence on the prediction. In this work, we employ Hybrid-CAM, a visualization method that combines global channel importance with high-resolution spatial activations, to generate more precise and faithful explanations of model behaviour. We show that even a comparatively simple convolutional neural network, when paired with an appropriate visualization technique, can yield substantially improved explainability. Quantitative insertion–deletion metrics, together with qualitative visual inspection, demonstrate that Hybrid-CAM produces more focused and reliable explanations, particularly at intermediate network layers, by concentrating on regions that truly drive the model’s predictions.