A-ADAPT: Adaptive Intracranial Artery Segmentation with Morphology-Guided Prompts and Difficulty-Aware Learning

Zhiwei Tan, Xin Wang, Meng Wang, Zixuan Liu, Yin Guo, Jiamin Xia, Niranjan Balu, Linda Shapiro, Chun Yuan, Mahmud Mossa-Basha
Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, PMLR 315:1509-1522, 2026.

Abstract

Accurate segmentation of intracranial arteries in CTA and MRA is essential for cerebrovascular analysis but remains challenging due to fine-scale artery morphology, modality-dependent appearance, and frequent structural discontinuities. Existing CNN or Transformer based models struggle to generalize across modalities, while SAM-based methods rely heavily on manually provided prompts and often fail to preserve thin or low-contrast arteries. We propose A-ADAPT, an adaptive intracranial artery segmentation framework that enhances SAM with modality-aware representation learning, automatic morphology-guided prompting, and difficulty-aware optimization. First, a Cross-Modality Task Adapter (CMTA) aligns CTA and MRA feature distributions while preserving shared vascular characteristics. The Frequency Adapter (FA) and the Tubular Morphology Adapter(TMA) work together to refine artery representation by enhancing structural detail and highlighting the continuity of tubular anatomy. To eliminate dependence on manual prompts, we introduce an Automatic Directional Morphology Prompt Encoder (AutoDM-Prompt), which generates artery-aware prompts directly from the input image. Additionally, a difficulty-aware loss dynamically upweights uncertain or discontinuity-prone regions, enabling the model to better recover small branches and reduce false positives. Experiments on CTA and MRA datasets show that A-ADAPT achieves higher accuracy, and better structural continuity compared to several state-of-the-art methods.

Cite this Paper

BibTeX
@InProceedings{pmlr-v315-tan26a, title = {A-ADAPT: Adaptive Intracranial Artery Segmentation with Morphology-Guided Prompts and Difficulty-Aware Learning}, author = {Tan, Zhiwei and Wang, Xin and Wang, Meng and Liu, Zixuan and Guo, Yin and Xia, Jiamin and Balu, Niranjan and Shapiro, Linda and Yuan, Chun and Mossa-Basha, Mahmud}, booktitle = {Proceedings of The 9th International Conference on Medical Imaging with Deep Learning}, pages = {1509--1522}, 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/tan26a/tan26a.pdf}, url = {https://proceedings.mlr.press/v315/tan26a.html}, abstract = {Accurate segmentation of intracranial arteries in CTA and MRA is essential for cerebrovascular analysis but remains challenging due to fine-scale artery morphology, modality-dependent appearance, and frequent structural discontinuities. Existing CNN or Transformer based models struggle to generalize across modalities, while SAM-based methods rely heavily on manually provided prompts and often fail to preserve thin or low-contrast arteries. We propose A-ADAPT, an adaptive intracranial artery segmentation framework that enhances SAM with modality-aware representation learning, automatic morphology-guided prompting, and difficulty-aware optimization. First, a Cross-Modality Task Adapter (CMTA) aligns CTA and MRA feature distributions while preserving shared vascular characteristics. The Frequency Adapter (FA) and the Tubular Morphology Adapter(TMA) work together to refine artery representation by enhancing structural detail and highlighting the continuity of tubular anatomy. To eliminate dependence on manual prompts, we introduce an Automatic Directional Morphology Prompt Encoder (AutoDM-Prompt), which generates artery-aware prompts directly from the input image. Additionally, a difficulty-aware loss dynamically upweights uncertain or discontinuity-prone regions, enabling the model to better recover small branches and reduce false positives. Experiments on CTA and MRA datasets show that A-ADAPT achieves higher accuracy, and better structural continuity compared to several state-of-the-art methods.} }
Endnote
%0 Conference Paper %T A-ADAPT: Adaptive Intracranial Artery Segmentation with Morphology-Guided Prompts and Difficulty-Aware Learning %A Zhiwei Tan %A Xin Wang %A Meng Wang %A Zixuan Liu %A Yin Guo %A Jiamin Xia %A Niranjan Balu %A Linda Shapiro %A Chun Yuan %A Mahmud Mossa-Basha %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-tan26a %I PMLR %P 1509--1522 %U https://proceedings.mlr.press/v315/tan26a.html %V 315 %X Accurate segmentation of intracranial arteries in CTA and MRA is essential for cerebrovascular analysis but remains challenging due to fine-scale artery morphology, modality-dependent appearance, and frequent structural discontinuities. Existing CNN or Transformer based models struggle to generalize across modalities, while SAM-based methods rely heavily on manually provided prompts and often fail to preserve thin or low-contrast arteries. We propose A-ADAPT, an adaptive intracranial artery segmentation framework that enhances SAM with modality-aware representation learning, automatic morphology-guided prompting, and difficulty-aware optimization. First, a Cross-Modality Task Adapter (CMTA) aligns CTA and MRA feature distributions while preserving shared vascular characteristics. The Frequency Adapter (FA) and the Tubular Morphology Adapter(TMA) work together to refine artery representation by enhancing structural detail and highlighting the continuity of tubular anatomy. To eliminate dependence on manual prompts, we introduce an Automatic Directional Morphology Prompt Encoder (AutoDM-Prompt), which generates artery-aware prompts directly from the input image. Additionally, a difficulty-aware loss dynamically upweights uncertain or discontinuity-prone regions, enabling the model to better recover small branches and reduce false positives. Experiments on CTA and MRA datasets show that A-ADAPT achieves higher accuracy, and better structural continuity compared to several state-of-the-art methods.
APA
Tan, Z., Wang, X., Wang, M., Liu, Z., Guo, Y., Xia, J., Balu, N., Shapiro, L., Yuan, C. & Mossa-Basha, M.. (2026). A-ADAPT: Adaptive Intracranial Artery Segmentation with Morphology-Guided Prompts and Difficulty-Aware Learning. Proceedings of The 9th International Conference on Medical Imaging with Deep Learning, in Proceedings of Machine Learning Research 315:1509-1522 Available from https://proceedings.mlr.press/v315/tan26a.html.

Related Material