Performance Evaluation and Ranking of Deep Learning Feature Extraction Models for Thyroid Cancer Diagnosis using D-CRITIC TOPSIS
Rohit Sharma, Gautam Kumar Mahanti, Ganapati Panda · 2023
The nodules in the thyroid region can be cancerous or non-cancerous, present even in healthy humans. Early diagnosis of thyroid cancer is helpful for prevention and treatment. Diagnosing thyroid cancer using traditional approaches is a hard-working task due to the considerable burden on the healthcare community. In this paper, we analyzed the performance of the AI models (Swin Transformer, Data Efficient Image Transformer, and Mixer Multi-layer Perceptron) to extract the features from the histopathological and ultrasound images. The Locally Linear Embedding (LLE) is used to reduce the dimensionality of the feature space. These transformed features are utilized for the training five classifiers models (Random Forest classifier, Naive Bayes, Logistic Regression, Support Vector Classifier, and k-nearest neighbors). There is a total of fifteen possible combination models tested using the 5-fold cross-validation technique, and three performance metrics are calculated. The recently proposed TOPSIS technique is used to benchmark the performance of all the models, and based on the TOPSIS scores, rank values are assigned. The distance correlation-based CRITIC (CRiteria Importance Through Intercriteria Correlation) is employed for the weight calculation of different performance criteria. The model with Swin Transformer as a feature extractor and Random forest as a classifier outperformed other models and achieved the highest TOPSIS score. The top-ranked model with a TOPSIS score of 1.0000 achieved an accuracy of 0.9335 and 0.8518 for thyroid ultrasound and histopathological images. The proposed models are simple and can be deployed on resource-constrained remote edge devices. With the help of IoT and 5G/6G communication technologies, either an ensemble model is created, or federated learning techniques can be utilized to transfer the weights in order to train the cloud-based global model.