Detection and Classification Method for Early-Stage Colorectal Cancer Using Dyadic Wavelet Packet Transform

Incorporating deep learning into computer-aided medical diagnosis has led to significant advancements. However, a major challenge remains in interpreting deep learning models, especially in identifying the features critical for diagnosis. This study proposes a method for the early detection of colorectal cancer. The method uses the dyadic wavelet packet transform, root mean square, and zero count metrics to extract meaningful features from endoscopic images. These features effectively capture variations in pixel values and delineate contours and patterns. A standout strength of the proposed method lies in its interpretability. The direct feature extraction process enhances both the method's transparency and the classification results' comprehensibility. Furthermore, by analyzing the mean absolute values of principal component scores, we provide insightful explanations for misclassification cases, offering a deeper understanding of the classification mechanics. We experimentally evaluated the performance of the proposed method against five deep learning models on three diverse datasets. The results showed that our method achieved a validation accuracy of 99.2% on the Saga University Hospital dataset, 97.4% on the Kvasir-SEG dataset, and 97.8% on the CVC-ClinicDB dataset. These scores surpassed all other models tested, with precision, recall, and F1-measure values consistently higher across datasets, highlighting the method's robustness and potential for reliable early colorectal cancer detection.