An integrated intelligent system for breast cancer detection at early stages using IR images and machine learning methods with explainability

• Main Authors: Aidossov, Nurduman, Zarikas, Vasilios, Zhao, Yong, Mashekova, Aigerim, Ng, Eddie Yin Kwee, Mukhmetov, Olzhas, Mirasbekov, Yerken, Omirbayev, Aldiyar
• Other Authors: School of Mechanical and Aerospace Engineering
• Format: Article
• Language: English
• Published: 2023
• Subjects: Engineering::Mechanical engineering; Thermography; Convolutional Neural Network
• Online Access: https://hdl.handle.net/10356/170074
• Institution: Nanyang Technological University

Breast cancer is the second most common cause of death among women. An early diagnosis is vital for reducing the fatality rate in the fight against breast cancer. Thermography could be suggested as a safe, non-invasive, non-contact supplementary method to diagnose breast cancer and can be the most promising method for breast self-examination as envisioned by the World Health Organization (WHO). Moreover, thermography could be combined with artificial intelligence and automated diagnostic methods towards a diagnosis with a negligible number of false positive or false negative results. In the current study, a novel intelligent integrated diagnosis system is proposed using IR thermal images with Convolutional Neural Networks and Bayesian Networks to achieve good diagnostic accuracy from a relatively small dataset of images and data. We demonstrate the juxtaposition of transfer learning models such as ResNet50 with the proposed combination of BNs with artificial neural network methods such as CNNs which provides a state-of-the-art expert system with explainability. The novelties of our methodology include: (i) the construction of a diagnostic tool with high accuracy from a small number of images for training; (ii) the features extracted from the images are found to be the appropriate ones leading to very good diagnosis; (iii) our expert model exhibits interpretability, i.e., one physician can understand which factors/features play critical roles for the diagnosis. The results of the study showed an accuracy that varies for the most successful models amongst four implemented approaches from approximately 91% to 93%, with a precision value of 91% to 95%, sensitivity from 91% to 92 %, and with specificity from 91% to 97%. In conclusion, we have achieved accurate diagnosis with understandability with the novel integrated approach.