Uncertainty Ordinal Multi-Instance Learning for Breast Cancer Diagnosis

Area of Science:

• Medical imaging analysis
• Machine learning
• Computer-aided diagnosis

Background:

• Breast cancer is a leading women's disease, necessitating early detection through mammography.
• Current methods often overlook normal mammogram classification and the ordinal relationships within instances.
• Distinguishing early benign lesions from normal tissue is challenging due to visual similarities.

Purpose of the Study:

• To develop an improved ordinal multi-instance learning (OMIL) method for enhanced breast cancer screening.
• To explicitly model ordinal class information for both bags and instances within bags.
• To address the limitations of existing OMIL approaches that ignore ordinal instance information.

Main Methods:

• Proposed a novel OMIL approach that designates a key instance within each bag.
• Introduced an ordinal minimum uncertainty loss function for iterative optimization of key instances.
• Applied the method to mammogram classification, considering three categories: normal, benign, and malignant.

Main Results:

• Achieved 52.021% accuracy, 61.471% sensitivity, 47.206% specificity, 57.895% precision, and 59.629% F1 score on the DDSM dataset.
• Demonstrated the effectiveness of the proposed ordinal instance-learning approach.
• The method successfully leverages ordinal information for improved classification.

Conclusions:

• The proposed OMIL method enhances early breast cancer screening by effectively classifying mammograms.
• Explicitly modeling ordinal relationships improves the performance of multi-instance learning in medical contexts.
• This approach offers a promising direction for computer-aided diagnosis in mammography.