Abstract

Machine learning-based breast cancer detection from mammograms is inherently affected by image degradation induced by imaging systems. Conventional algorithms mainly target enhancement of classification performance in an ideal imaging scenario; little consideration is generally given to the impact of imaging degradations on robustness and interpretability of the model. In this paper, a robust approach for systematically investigating the effect of essential imaging system features, such as spatial resolution, image blur, and additive Gaussian noise concerning patch-based mammogram cancer detection has been introduced. A convolutional neural network combined with a circular attention mechanism is used to extract discriminative features from fixed-size mammogram patches and provide an interpretable visualization. The model is trained on high-quality, non-degraded patches and is evaluated on degraded patches to investigate the effect of different imaging factors. The proposed system incorporates patch-based learning, degradation modeling, quantitative performance evaluation, lesion localization analysis, and attention-based interpretability within the developed framework to formulate an integrated evaluation of the model functionality under realistic acquisition environment. Experimental results indicated that spatial resolution was the most crucial factor for detection performance, and even modest resolution degradation brought a major drop in cancer sensitivity, lesion localization performance, and attention-map quality. The investigations also showed that the model is relatively robust for mild Gaussian blur and moderate Gaussian noise. It was noted that blur primarily affects edges and texture of the processed mammograms whereas noise mostly diminishes classification specificity by inducing false positive outputs and preserving lesion detectability. Attention visualizations showed that imaging degradations decrease feature localization and diminish the ability of the network to focus on areas important for diagnosis. The research emphasizes the importance of image quality for reliable mammogram analysis using machine learning. Further, it shows that the traditional accuracy metrics results in lower diagnostic integrity. Thus, the proposed framework results in the development of robust, interpretable, and clinically implementable computer-aided diagnosis systems for breast cancer detection particularly for rural clinics where access to high-end imaging systems is generally limited.

Statistics