Enhancing medical image classification with cross-dimensional transfer learning using deep learning

Breast cancer is a major health issue, and effective treatment depends on a prompt diagnosis. Particularly mammography is important in the detection of breast cancer. Deep learning algorithms have shown promise in analyzing medical images, but their performance heavily relies on large labeled datasets, which are often limited in the context of breast cancer. In spite of the lack of labeled data, this study suggests a unique method called Cross-Dimensional Transfer Learning to increase the precision of cancer identification using deep learning. The method utilizes multiple imaging modalities, such as mammography and ultrasound, to leverage the complementary information and transfer knowledge learned from one modality to enhance classification performance on another. The proposed work consists of following three phases: Pretraining on Diverse Data, Modality-Specific Fine-Tuning and Cross-Dimensional Transfer Learning. A deep learning model is pretrained on a diverse dataset that includes breast cancer images from different modalities. This phase enables the model to learn general features and representations applicable across various imaging modalities. After pretraining, the model is perfected using labeled data specific to each modality. This process enables the model to adapt its learned features to the exclusive features of each imaging modality, improving its ability to capture modality-specific patterns related to breast cancer. Once modality-specific fine-tuning is complete, knowledge acquired from one modality is transferred to another by leveraging shared representations between the imaging modalities. This transfer of knowledge enhances the classification performance on the target modality, particularly when labeled data is limited for that modality.