DIAGNOSIS OF THYROID CANCER USING BOBCAT OPTIMIZATION ALGORITHM (BOA) AND ENHANCED TWO STAGE ATTENTION WITH LONG SHORT-TERM MEMORY (ETSA-LSTM)

Abstract

The malignant (cancerous) cells develop in the thyroid gland's tissues, is known as thyroid cancer (TC). For these conditions to be effectively treated and for patient care to be provided, an accurate and quick diagnosis is essential. Thyroidectomy is still the principal therapeutic option, despite significant efforts to improve diagnosis. A precise preoperative diagnosis may not always be ensured by the current human evaluation of Thyroid Nodule (TN) malignancy, because this TN maligancy is prone to errors. Medical (DA) Data Analysis (MDA) problems can be easily solved with the use of Data Mining (DM) algorithms. With several techniques for classification, clustering, association, etc., DM offers significant assistance with thyroid datasets. Deep Learning (DL) techniques to predict and identify TC, their development and application present a number of difficulties. The suggested work consists of five primary steps: Data Collection (DC), Data Pre-processing, Feature Selection (FS), Data Classification, and Performance Evaluation. Initially, the Kaggle online repository is used to gather the DC, TC risk prediction dataset. This dataset mimics real-world TC risk factors and includes 212,691 records with 23 features. Then, issues with Data Encoding (DE), Data Resampling (DR), Data Normalization (DN), Data Imputation (DI) and handling missing data are addressed by using data pre-processing techniques. DN is carried out using Min-Max Normalisation (MMN) or Min-Max Scaling (MMS). To extract pertinent feature information, the Z-score- (ZS) based Boruta-SHapley Additive exPlanations (BorutaSHAP) attribute importance technique, is used by the Feature importance (FI). SMOTE-EkNN is a data imbalance problem-solving technique that uses Edited k-Nearest Neighbors (EkNN) and Synthetic Minority Over-sampling Technique (SMOTE) to remove noise. Thirdly Feature Selection (FS), Bobcat Optimization Algorithm (BOA) is used to choose appropriate features. For TC detection, the most pertinent features (an optimal reduced feature subset) are chosen using BOA. The Search Space (SS) can be effectively explored and exploited by BOA, a Nature-Inspired Algorithm (NIA). Next, the data is classified using Enhanced Two Stage Attention with Long Short-Term Memory (ETSA-LSTM) for predicting the TN malignancy. Accuracy (ACC), Specificity (SP), Precision (PR), Recall (R) or Sensitivity (SE), and the Area Under Receiver Operating Characteristic (AUROC) curve are the final metrics used to evaluate efficiency. According to the study of the results, the suggested model outperformed the other methods currently in use in terms of ACC .