Beyond Classification: Elaborating Network Predictions for Better Weakly Supervised Quantization

Chih-Chieh Chen, Chang-Fu Kuo
Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, PMLR 315:1-20, 2026.

Abstract

For clinical applications, more detailed information such as specific locations and the region of interest (ROI) volumes is preferred. However, most of the time only classification annotations are available. Class Activation Mapping (CAM) and its variants are the most commonly used techniques for weakly supervised localization tasks. In this study, we assessed both traditional and modern network architectures regarding classification accuracy and CAM visualization. Although all networks achieved high AUROC scores and their heatmaps closely corresponded to pathology locations, we observed that the heatmaps were influenced by the particular network architectures and pretrained weights used. Additionally, current models produce heatmaps from small latent spaces (e.g. $16 \times 16$), which limits the precision of these heatmaps for further detailed analysis. Based on the observations mentioned above, we designed a UNet-style architecture that utilizes pretrained classification networks as the encoder and produces heatmaps within a latent space of size $128 \times 128$. We observed that the generated heatmaps are more detailed and suitable for weakly supervised segmentation. We validated the effectiveness of our approach using the intracerebral hemorrhage (ICH) dataset.

Cite this Paper

BibTeX
@InProceedings{pmlr-v315-chen26a, title = {Beyond Classification: Elaborating Network Predictions for Better Weakly Supervised Quantization}, author = {Chen, Chih-Chieh and Kuo, Chang-Fu}, booktitle = {Proceedings of The 9th International Conference on Medical Imaging with Deep Learning}, pages = {1--20}, year = {2026}, editor = {Huo, Yuankai and Gao, Mingchen and Kuo, Chang-Fu and Jin, Yueming and Deng, Ruining}, volume = {315}, series = {Proceedings of Machine Learning Research}, month = {08--10 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v315/main/assets/chen26a/chen26a.pdf}, url = {https://proceedings.mlr.press/v315/chen26a.html}, abstract = {For clinical applications, more detailed information such as specific locations and the region of interest (ROI) volumes is preferred. However, most of the time only classification annotations are available. Class Activation Mapping (CAM) and its variants are the most commonly used techniques for weakly supervised localization tasks. In this study, we assessed both traditional and modern network architectures regarding classification accuracy and CAM visualization. Although all networks achieved high AUROC scores and their heatmaps closely corresponded to pathology locations, we observed that the heatmaps were influenced by the particular network architectures and pretrained weights used. Additionally, current models produce heatmaps from small latent spaces (e.g. $16 \times 16$), which limits the precision of these heatmaps for further detailed analysis. Based on the observations mentioned above, we designed a UNet-style architecture that utilizes pretrained classification networks as the encoder and produces heatmaps within a latent space of size $128 \times 128$. We observed that the generated heatmaps are more detailed and suitable for weakly supervised segmentation. We validated the effectiveness of our approach using the intracerebral hemorrhage (ICH) dataset.} }
Endnote
%0 Conference Paper %T Beyond Classification: Elaborating Network Predictions for Better Weakly Supervised Quantization %A Chih-Chieh Chen %A Chang-Fu Kuo %B Proceedings of The 9th International Conference on Medical Imaging with Deep Learning %C Proceedings of Machine Learning Research %D 2026 %E Yuankai Huo %E Mingchen Gao %E Chang-Fu Kuo %E Yueming Jin %E Ruining Deng %F pmlr-v315-chen26a %I PMLR %P 1--20 %U https://proceedings.mlr.press/v315/chen26a.html %V 315 %X For clinical applications, more detailed information such as specific locations and the region of interest (ROI) volumes is preferred. However, most of the time only classification annotations are available. Class Activation Mapping (CAM) and its variants are the most commonly used techniques for weakly supervised localization tasks. In this study, we assessed both traditional and modern network architectures regarding classification accuracy and CAM visualization. Although all networks achieved high AUROC scores and their heatmaps closely corresponded to pathology locations, we observed that the heatmaps were influenced by the particular network architectures and pretrained weights used. Additionally, current models produce heatmaps from small latent spaces (e.g. $16 \times 16$), which limits the precision of these heatmaps for further detailed analysis. Based on the observations mentioned above, we designed a UNet-style architecture that utilizes pretrained classification networks as the encoder and produces heatmaps within a latent space of size $128 \times 128$. We observed that the generated heatmaps are more detailed and suitable for weakly supervised segmentation. We validated the effectiveness of our approach using the intracerebral hemorrhage (ICH) dataset.
APA
Chen, C. & Kuo, C.. (2026). Beyond Classification: Elaborating Network Predictions for Better Weakly Supervised Quantization. Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 315:1-20 Available from https://proceedings.mlr.press/v315/chen26a.html.

Related Material