SkinSight: A Novel Implementation of a Convolutional Neural Network to Recognize Skin Diseases
CSEF · 2019 Computational Systems & Analysis (Senior Division Only)
Overview
Objectives Only 35% of skin disease biopsies are performed by dermatologists, and non dermatologists only correctly diagnose skin diseases with 23.9% accuracy. Many methods have tried to solve this issue, but most are computationally heavy, while algorithms that can segment lesions have not been used for diagnosis. This project aims to design a computationally light and effective convolutional neural network (CNN) to diagnose skin lesions while also evaluating two common deep learning based approaches, Mask R-CNN and FCN, for multiclass segmentation against results from the ISIC 2018 Challenge: Skin Lesion Analysis Towards Melanoma Detection. Methods Our physical materials included a laptop with Anaconda installed on it. The dataset was provided by DermnetNZ, but since there are only 4374 images, we used data augmentation to make it larger. To train Mask R-CNN and FCN, we produced a dataset of 520 images with pixelwise masks identifying the lesions. Then, we used code for Mask R-CNN and FCN from Matterport and Sagippel s Github, respectively, and for transfer learning, we used the Tensorflow Image Retraining Tutorial. For our CNN, we used transfer learning, where we took an existing CNN, Inception-v3, and added a classification layer. To improve accuracy, we grouped diseases by physical traits. For Mask R-CNN, we wrote a new supervision and dataset loading file and trained it for 60 epochs. For FCN, it trained for approximately 40 epochs. Results We found that as the size of each group for our transfer learning decreased, accuracy improved, but human error increased when choosing the correct group. As the number of training steps increased, the accuracy increased. For the segmentation techniques, the Mask R-CNN achieved 0.9053 mean average precision and a 0.7532 Jaccard Index while the FCN achieved a 0.7241 Jaccard Index. Conclusions Having 15 diseases per group and 16000 training steps offered the best cost/efficiency tradeoff for its 86.5% balanced accuracy and 0.85 average F1 score. FCN was less effective than Mask R-CNN because it introduced interference into its predictions, leading to a lower Jaccard Index. The transfer learning and segmentation methods we tested in this study achieved comparable results to studies in the ISIC 2018 Challenge which achieved 85.3% balanced accuracy and 0.813 Jaccard Index, proving the feasibility of both transfer learning and multiclass segmentation.
Summary statement
We implemented an effective and efficient transfer learning approach to skin lesion diagnostics, and we also proved that multiclass segmentation of skin lesions is feasible, with the best method of doing so being Mask R-CNN.
Help received
None. We worked on the project by ourselves.
Competition history
- CSEF 2019
Resources
Related projects
CSEF · 2018
skinCAM: An Effective Solution towards Skin Disease Diagnosis via Novel Deep Learning Algorithm
CSEF · 2019
Developing a Novel, Accurate, and Rapid Machine Learning Based Skin Disease Diagnosis Algorithm and Mobile Application
CSEF · 2014
Accurate Detection of Skin Cancer Using Multi Stage Neural Networks
ISEF · 2017
Scin: An Application to Detect and Diagnose Plant and Skin Diseases Utilizing a Feed Foward Convolutional Neural Network
ISEF · 2017
Utilizing Machine Learning Techniques to Identify Cancerous Skin Lesions
ISEF · 2024
Rapid Screening for Major Skin Cancers with M-SCAHN: Multimodal Hierarchical CNN-Transformer Hybrid Networks with Advanced Interpretability, Lesion Evolution Tracking, Trait Identification, and Color Constancy for Improved Generalization in Diverse Populations
ISEF · 2017
A Novel Machine Learning Approach Using Convolutional Neural Networks to Identify Melanoma
ISEF · 2018
Detection of Melanoma via Deep Learning
Closest projects by meaning, across every fair and year in the corpus.
Browse more like this
Source: California Science & Engineering Fair public projects