Utilizing Densely Connected Convolutional Network Architectures with Effective Feature Reuse for the Enhanced Early Detection of Pancreatic Neuroendocrine Malignancies
Nirvaan Duggirala, Saketh Kollipara · 2025
Pancreatic Neuroendocrine Tumors (PNETs) pose an alarmingly dangerous effect on society being one of the most lately diagnosed or misdiagnosed forms of cancer, with a lack of methods that don’t focus on early diagnosis, leading to the mortality rate being extremely high at 77% after metastasis. Deep learning can make a large contribution to the process of classification and predictions of malignant PNETs in comparison with normal cases, and has been proven to be a strong technique with other cancers. In this study, we propose a novel framework that utilizes Densely Connected Convolutional Networks (DenseNets) for the enhancement of the current lack of early detection of PNETs, with a specific focus on early metastasis to liver lesions. Our methodology involves the implementation and fine-tuning of DenseNets, and specifically DenseNet-121, leveraging the dense connectivity between all the layers in the network to allow for more comprehensive feature extraction and retention, as well as better feature reuse and reduced training parameter usage, conserving critical computational resources in the process. Quantitatively, our model performs exceptionally with an overall accuracy of 99.95%, a precision of 99.98%, and an AUC score of 99.999975%, suggesting a possible path towards a clinical impact associated with early tumor identification for cancer diagnosis.