Discriminative Dictionary Learning Using Penalized Rank-1 Approximation for Breast Cancer Classification With Imbalanced Dataset

In histopathological image analysis, the feature extraction task for classification proves to be demanding. This difficulty arises from the assortment of histological features appropriate for individual problems and the intricate presence of diverse geometric structures. The method proposed in this study leverages dictionary learning and sparse coding techniques to create priors tailored to specific targets, which is essential for classification purposes. Our approach introduces Penalized Sequential Discriminative Dictionary Learning (PSDDL), designed to integrate histopathological image features by acquiring structured, class-specific dictionaries. Initially, PSDDL constructs a dictionary from the input data, incorporating label information for each class. Subsequently, the proposed algorithm introduces a penalty and regularization term to amplify the efficacy of the acquired dictionary. Furthermore, the proposed method also tackles the class imbalance in the dataset by leveraging dictionary learning. Mainly, under-sampling is performed on the dataset, and the number of samples of fewer classes is kept for all categories and passed to the dictionary learning algorithm. Extensive experimental results highlight a notable enhancement in classification performance. The proposed structured discriminative dictionary learning technique consistently produces improved accuracies compared to other contemporary methods for classifying the BreakHis breast cancer dataset. Integrating class-specific information into the process of dictionary learning paves the way for enriching the interpretive capabilities of machine learning models while delving deeper into our comprehension of the complex and intricate structures inherent in biological tissues.