Automatic Blob Detection Method for Cancerous Lesions in Unsupervised Breast Histology Images

Area of Science:

• Medical Image Analysis
• Computational Pathology
• Deep Learning in Oncology

Background:

• Early breast cancer lesion detection is crucial but challenging due to tissue variability and limitations of conventional computer-aided diagnosis (CAD).
• Traditional methods struggle with large images, overlapping objects, and accurately defining lesion boundaries, hindering efficient analysis.

Purpose of the Study:

• To develop and evaluate a deep learning-based blob detection technique for automatically identifying hidden cancerous lesions in unsupervised human breast histology images.
• To enhance the accuracy and efficiency of breast cancer lesion detection by overcoming the limitations of conventional methods.

Main Methods:

• Data augmentation and stain normalization were applied to histology images.
• Morphology operations (erosion, dilation, opening, distance transform) and connected components analysis were used for image enhancement and segmentation.
• A deep learning recurrent neural network (RNN) model integrated with active contours method performed blob detection to identify lesions and their edges.

Main Results:

• The proposed deep learning blob detection method was evaluated on a dataset of 27,249 unsupervised, augmented human breast cancer histology images.
• The technique demonstrated a significant evaluation result with a 98.82% F1 accuracy score.
• The approach effectively detected hidden and inaccessible cancerous lesions and their boundaries.

Conclusions:

• The developed deep learning technique offers an effective solution for the early detection of breast cancer lesions in histology images.
• By combining connected components and active contours methods, the approach addresses limitations in blob detection, improving lesion identification.
• This method shows promise for improving the accuracy and efficiency of pathological analysis in breast cancer diagnosis.