Tumor Bud Classification in Colorectal Cancer Using Attention-Based Deep Multiple Instance Learning and Domain-Specific Foundation Models DOI Open Access
Mesut Şeker, Muhammad Khurram Khan, Wei Chen

и другие.

Cancers, Год журнала: 2025, Номер 17(7), С. 1245 - 1245

Опубликована: Апрель 7, 2025

Identifying tumor budding (TB) in colorectal cancer (CRC) is vital for better prognostic assessment as it may signify the initial stage of metastasis. Despite its importance, TB detection remains challenging due to subjectivity manual evaluations. TBs difficult, especially at high magnification levels, leading inconsistencies prognosis. To address these issues, we propose an automated system classification using deep learning. We trained a learning model identify through weakly supervised by aggregating positive and negative bags from invasive front. assessed various foundation models feature extraction compared their performance. Attention heatmaps generated attention-based multi-instance (ABMIL) were analyzed verify alignment with TBs, providing insights into interpretability features. The dataset includes 29 WSIs training 70 whole slide images (WSIs) hold-out test set. In six-fold cross-validation, Phikon-v2 achieved highest average AUC (0.984 ± 0.003), precision (0.876 0.004), recall (0.947 0.009). again (0.979) (0.980) on external Moreover, rate (0.910) was still higher than that UNI's (0.879). UNI exhibited balanced performance set, 0.960 0.968. CtransPath showed strong set (0.947) but had slightly lower (0.911). proposed technique enhances accuracy assessment, offering potential applications CRC other types.

Язык: Английский

Tumor Bud Classification in Colorectal Cancer Using Attention-Based Deep Multiple Instance Learning and Domain-Specific Foundation Models DOI Open Access
Mesut Şeker, Muhammad Khurram Khan, Wei Chen

и другие.

Cancers, Год журнала: 2025, Номер 17(7), С. 1245 - 1245

Опубликована: Апрель 7, 2025

Identifying tumor budding (TB) in colorectal cancer (CRC) is vital for better prognostic assessment as it may signify the initial stage of metastasis. Despite its importance, TB detection remains challenging due to subjectivity manual evaluations. TBs difficult, especially at high magnification levels, leading inconsistencies prognosis. To address these issues, we propose an automated system classification using deep learning. We trained a learning model identify through weakly supervised by aggregating positive and negative bags from invasive front. assessed various foundation models feature extraction compared their performance. Attention heatmaps generated attention-based multi-instance (ABMIL) were analyzed verify alignment with TBs, providing insights into interpretability features. The dataset includes 29 WSIs training 70 whole slide images (WSIs) hold-out test set. In six-fold cross-validation, Phikon-v2 achieved highest average AUC (0.984 ± 0.003), precision (0.876 0.004), recall (0.947 0.009). again (0.979) (0.980) on external Moreover, rate (0.910) was still higher than that UNI's (0.879). UNI exhibited balanced performance set, 0.960 0.968. CtransPath showed strong set (0.947) but had slightly lower (0.911). proposed technique enhances accuracy assessment, offering potential applications CRC other types.

Язык: Английский

Процитировано

0