Abstract:
Breast cancer is a major cause of cancer mortality in women and therefore early
and precise detection is essential in treatment. The traditional diagnostic
techniques like mammography and histopathology are still popular and are
constrained by the variability of the observer and sensitivity in complicated
cases. Convolutional Neural Networks (CNNs) Deep learning has become a
potent remedy, being able to extract features automatically with high accuracy
on classification. Nevertheless, there are still issues in small data sets, binary
only classification, and low interpretability. This paper will suggest a hybrid
architecture combining CNNs and Capsule Network to enhance histopathologybased breast cancer detection. The results have been assessed using 6,000
images preprocessed and tested on various architectures, which are
EfficientNetB0, InceptionV3, ResNet variants, and MobileNet V2. Among them,
CapsuleNet model had the highest accurate, precise, recall and F1-score of 96.44
that surpassed any baselines tested. Gradient-weighted Class Activation
Mapping (Grad-CAM) was also used to make the results more interpretable,
giving heatmaps that show which parts of the image affected the models in the
results, and this particular aspect should give the results more transparency in
clinical practice. To be implemented practically, the optimized CapsuleNet model
was translated to a mobile-based application with the help of TensorFlow Lite.
This provides accessibility in resource constrained healthcare settings and
proves the possibility of integrating in real world settings. Since it integrates
good diagnostics and explainability and portability, the proposed solution can be
useful in academic studies and clinical practice, helping to diagnose diseases
earlier, minimize false diagnosis, and lead to better patient outcomes.
