DEVELOPMENT OF FRAMEWORK FOR THERMAL PATTERNS ANALYSIS FOR DIAGNOSIS OF DISEASES

Abstract

Attitude towards quality of life in the current century is unhealthy, result- ing in several diseases. Consequently, life expectancy has significantly declined since 2000. Various medical imaging procedures like X-rays, Magnetic Resonance Imaging, Ultra-Sonography, etc., in combination with pathological tests form the basis for all medical conclusions. However, these imaging procedures are inva- sive in nature, require specialized radiologists, and not accessible to the populace. Hence, Thermography-based technology is proposed in this thesis for screening and diagnosing abnormality/inflammation in human body. Medical thermal imag- ing is unique in its potential to demonstrate the physiological change and metabolic processes in a human body through thermal patterns radiated by it. Thermal imag- ing technology accompanied with robust and automated computational-aided diag- nostic systems can be deployed at public gathering locations (like malls, hospitals, etc.) and on smartphones to timely detect and warn about potential health issues of their body caused by inflammation. Thermography-based cameras have shown their efficacy during the recently seen pandemic time in screening the physical behavior of the human body and re- porting unhealthy individuals instantly. Furthermore, improper use of Thermography- based technology may result in wrong temperature readings due to their low res- olution. This thesis have examined its potential in conjunction with Machine and Deep Learning techniques for identifying inflammation in the human body as a classification problem. We have proposed two statistical models aimed at differen- tiating normal and abnormal thermal patterns in medical thermal imaging. The first model utilizes a novel of asymmetry-based features extracted from three publicly available datasets, focusing on detecting abnormalities related to breast cancer, diabetes, and thyroid disorders. The second model systematically evaluates the performance of this proposed feature set in comparison with eight state-of-the-art feature extraction techniques, establishing a standardized methodology for analyz- ing medical thermal images. To ensure unbiased evaluation, a two-level sampling strategy was employed to address dataset imbalance, and cross-validation tech- niques. A lightweight deep learning-based classification model designed for mo- bile deployment is presented to detect and characterize abnormalities caused by inflammation in human thermal images. In the context of medical thermal image segmentation, we present two novel density-based modified PC FS techniques, - DSIFC-PcFS and DSIMFC-PcFS, for vi segmenting inflamed regions in abnormal thermal images. The limitations of ex- isting segmentation models include reliance on private datasets with limited sam- ples, subjectivity in ground truth generation, and sensitivity to parameter selection, thereby the proposed methods aim to improve robustness, accuracy, and reliability in medical thermal image analysis. The first model (DSIFC-PC FS) uses a density- based heuristic to automatically determine cluster centers and membership values. Furthermore, spatial information is integrated into the model to reduce sensitiv- ity to noise and preserve fine image structures without requiring prior smoothing. The second model (DSIMFC-PC FS) further refines clustering by incorporating modified Renyi’s entropy to improve segmentation accuracy and optimize cluster partitions. The models are optimized using Lagrangian methods and validated on thermal imaging datasets covering diabetic foot, breast cancer, and thyroid disor- ders, with and without artificial noise (Gaussian, Salt & Pepper, and Mixed Noise). We evaluated the performance of all developed methods using publicly avail- able medical thermal imaging datasets through both visual and quantitative as- sessments. Additionally, we compared their performance with well-established and state-of-the-art algorithms. Statistical analysis was conducted using paired T- Test and Friedman Test, demonstrating the superiority of the proposed methods over existing state-of-the-art algorithms in classification and segmentation tasks. Overall the findings establish a strong foundation for future research in automated thermography-based diagnostic systems, particularly for early disease detection in real-world clinical settings.