Performance Improvement of Lung Cancer Multiclass Classification on Transformer Convolution Neural Networks using Histopathalogical Images

Globally, lung cancer ranks among the top causes of death for both men and women. The field of early detection and identification of lung cancer depends heavily on the prompt and precise classification of this disease. Historically, human error has resulted in delays in diagnosis and the possibility of misdiagnosis due to dependence on traditional, subjective approaches. Pathologists frequently use biopsy specimens images from potentially sick lung areas to determine the kind and likelihood of malignancy. But this process is also laborious and occasionally flawed. The growing amount of medical imaging information puts radiologists under a lot of strain, hence a more automated and efficient method is required. One of the current benchmarks for automated lung cancer detection systems is represented by Deep Neural Networks that use histopathology pictures as an input. The proposed method introduces an efficient and robust Transformer Convolutional Neural Network (TCNN) model tailored specifically for lung nodule classification. The proposed study has to improve the speed, accuracy of lung cancer classification, allowing for early diagnosis and accurate classification of different forms of disease. The proposed technique enhances the model resistance to variations in external illumination by focusing on its Hue-Saturation-Value colormap layers that is associated with colour representation. A series of convolutional neural network layers is utilised during the Efficient Net neural network for classification, and spatial converter layer were added to build the model substantially consistent on input information. More than 15,000 images are utilised for training and validation, and they show three different forms of lung cancer: The three types of lung tissues that have been investigated include benign tissue, squamous cell carcinoma, and adenocarcinoma. We assess the model using performance evaluation, utilising many measures, such as the confusion matrix, the F1 score, and accuracy. Finally, 98.61% training and 98.61% validation accuracy were obtained using the proposed approach.