Optimization of Breast Cancer Classification with Octave Bands Analysis

Recent advancements in medical imaging technologies with a particular focus on the integration of artificial intelligence in image analysis, show potential in tackling clinical challenges related to the detection of breast cancer, evaluating treatment responses, and monitoring the progression of the disease. In this paper, a new algorithm based on bands that cross the lesion (three verticals, and three horizontals), with widths of 2, 4, and 8 pixels to classify breast lesions is proposed. A new octave band analysis is proposed to optimize the model's feature extraction. Thus, each selected band is split into twelve sub-bands, and seventy-two features are obtained. The vector features dimensionality is reduced based on the features' meaningfulness. To overcome the various machine learning models that fail to accurately classify these images due to their complex and diverse nature, several Automated Machine Learning libraries (AutoML) with PyCaret are used. They tune the hyperparameters for each selected classifier. For each selected band only eight features feed an AutoML PyCaret algorithm. In the case of the band having a width of 4 pixels, the binary classification results provide an accuracy of 0.812, an Area Under the Curve (AUC) of 0.844, and an F1-score of 0.867. The results are validated by the bagging method. It provided for the test dataset an accuracy of 0.813.