Early Detection of Acromegaly Using a Novel Convolutional Neural Network

Overview

Acromegaly occurs when the pituitary gland produces too much somatropin, causing the liver to release excessive amounts of IGF-1, leading to the abnormal growth of the hands, feet, and face. Acromegaly is difficult to diagnose and can lead to serious, sometimes even life-threatening, health problems such as Type II diabetes and heart disease. The early detection of Acromegaly reduces potential health complications and the risk of death. Deep learning assisted early detection has now been proven feasible according to latest research and the prevalent success of Transfer Learning provides a potential path of non-computationally intensive detection. In this study, a dataset containing roughly 20 images were used to train a Convolutional Neural Network with Transfer learning that utilized ResNet-18 to mitigate the low dataset size. Firstly, Acromegaly and Non-Acromegaly images were placed into separate datasets and were further separated in a 70/30 training- validation split. This was run through the model, achieving a 65.62% validation accuracy over 25 epochs. This was paired with high training-validation loss values, 0.7/1.5 respectively, past epoch 25. T o improve these losses, pairs of the same person were used to mitigate data imbalance within the datasets occurring from multiple same patient images. The datasets were comprised of Acromegaly/Non-Acromegaly (A/N) pairs and Non Acromegaly/ Non-Acromegaly (N/N) pairs and each pair was fed through a custom data loader to then be fed through two Resnet-18 models, which were able to train on the differences between normal (N/N) and abnormal (A/N) growth. This led to a 9.9% validation increase as well.

Competition history

• JSHS 2023 Category not listed

Resources

• National JSHS abstract book ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.