Comparative analysis of Deep CNN Models for Pancreatic Cancer with Adam Optimizer
Indroneil Sinha Roy, Kalyan Acharjya, Arun Kumar Singh · 2024
In the field of cancer research, the possibility of increasing the effectiveness of diagnostic tools is based on the classification of biomedical data correctly. That results in improving patient care and outcomes. The best-fitting deep learning models must be identified to ensure the highest performance possible regarding a multitude of datasets. However, although the wide variety of deep learning models available for diagnostic purposes is diverse, their evaluations concerning many datasets remain limited. In this research, seven deep-learning model architectures were examined and explored using the Clinical Proteomic Tumor Analysis Consortium - Pancreatic Ductal Ade-nocarcinoma (CPTAC-PDA) and Kaggle datasets. The task is to identify the best models for every dataset and ordinary capacity of generalization with Adam optimizer. The analysis were done on comparing test accuracy, precision, recall and F1-score of every model in each of the datasets. All tested models should be trained and tested in similar conditions; the comparison of performance metrics will help to understand the strong and weak sides of each technique. The computations revealed the performance by ResNet50 as the best in CPTAC-PDA and showed a higher capacity of generalization by InceptionV3 and InceptionResNetV2 in a Kaggle dataset. The results suggest that the performance of DenseNet121 and MobileNetV2 is not sufficiently sustainable. The proposed models, particularly ResNet50, achieved superior performance, with ResNet50 attaining a 96.08% accuracy and 0.9744 F1-score, surpassing reference models. VGG16 and InceptionV3 also showed strong results, further validated by improved ROC curves. Findings aid in creating robust diagnostic tools.