Domain and Histopathology Adaptations-Based Classification for Malignancy Grading System.

Area of Science:

• Oncology
• Computer Science
• Medical Imaging

Background:

• Manual grading of breast cancer from histopathology slides is subjective and resource-intensive.
• Accurate proliferation rate quantification is crucial for effective breast cancer treatment planning.
• Convolutional neural networks (CNNs) offer potential for objective and efficient image analysis in pathology.

Purpose of the Study:

• To develop a novel, automated scheme for grading breast cancer malignancy in invasive ductal carcinoma.
• To enhance clinicians' diagnostic capabilities through a computer-based observer.
• To create a patchless grading system for both magnification-dependent and independent classifications.

Main Methods:

• Employed multistage transfer learning with domain and histopathologic transformations for classification.
• Utilized pretrained models (InceptionResNetV2, InceptionV3, NASNet-Large, ResNet50, ResNet101, VGG19, Xception) for domain adaptation on the BreaKHis dataset (×40 magnification).
• Selected InceptionV3 and Xception for categorizing the Databiox database into grades 1, 2, or 3, leveraging their pretrained weights.

Main Results:

• Achieved outstanding performance in magnification-dependent classification with overall accuracy ranging from 90.17% ± 3.08% to 97.67% ± 1.09%.
• Reported F1 scores between 0.9013 and 0.9760 for magnification-dependent classification.
• Demonstrated the effectiveness of the proposed automated grading system on different datasets and magnification levels.

Conclusions:

• The proposed automated breast cancer grading system shows high accuracy and reliability.
• This approach can assist in developing robust breast cancer grading systems for clinical applications.
• The patchless, CNN-based method offers a promising alternative to manual histopathological evaluation.