Integrating Adam Optimizer to Enhance Efficiency of Transfer Learning Model for Diagnosing Cancers

Abstract

Despite huge advancements in medical fields, diseases like cancer continue to plague people since we are still prone to them. The main purpose is to design a critical analysis of cancers, including breast, pancreatic, colon, skin, lung, and other cancers from the year 1990 to 2019, to identify some of the rising factors of both lung and colon cancers that cause unnecessary deaths. Thus, early detection on the other hand greatly improves the chances of survival rates in humans. At that place, Transfer Learning (TL) techniques have made significant improvements these days.

Material and Methods: We herein implemented the exploratory data analysis (EDA) technique to analyze 29 types of cancer deaths worldwide since 1990th century. Between them, we have emphasized lung cancer (LC) and colon cancers (CC) for our study and taken 25,000 histopathology image data from the publicly available Kaggle repository. Further, we proposed a novel methodology using the EfficientNetB7 model defining each step to classify five types of lung and colon tissues (two benign and three malignant) from the histopathological images. Although, an algorithm has been designed that clearly describes the workflow of our proposed method. This experiment has been done through input_shape=(X, Y, 3) on the ‘ImageNet’ dataset, Adam optimizer, and accuracy metrics using Python 3.8.8 software on Jupyter 6.4.3 environment.

Results: We evaluated the model progress using 50 epochs during the training phase, which resulted in more than 98% accuracy and less than 2% loss in both the training/ validation phases. Besides that, we designed a classification report that describes the performance of our model, where we achieved more than 98.07% accuracy score, 98% precision score, 98% of recall score, and 98% of f1-score. These results signify that implementing an optimization technique on the EfficientNetB7 model improves the overall performance as compared to 93.19% for LC 1 and 93.91% for CC 2 for the RestNet50 model.

Conclusion: A novelty of this research specifies the methodology to predict LC and CC by applying the EfficientNetB7 model to the cancer dataset which might be beneficial for both doctors as well as healthcare firms.