Hair Removal Algorithm and Transfer Learning
Dataset: https://lnkd.in/gNam-NpD
🔍 Project Insights: In this Research, I explored cutting-edge techniques in deep learning to advance the accuracy and efficiency of skin cancer diagnosis. Leveraging EfficientNetV2's strategic architecture adjustments and Vision Transformers self-attention mechanisms, I aimed to capture intricate patterns in skin images, surpassing limitations of traditional methods.
🛠️ Technical Approach: Harnessing pre-trained models from TensorFlow & Keras, I’ve have used 15 EfficientNet and EfficientNetV2 functions validate their performance on training images. Additionally, I customized Vision Transformer structures especially ViT8x8, adapting them for detailed feature extraction from 32x32 pixel images. These modifications aimed to enhance diagnostic precision without sacrificing computational efficiency.
📊 Model Evaluation: Through rigorous training sessions and meticulous testing phases, I assessed the performance of various models. Notably, EfficientNetV2 B2 emerged as a frontrunner, achieving a remarkable 98.5% accuracy in internal testing. External testing against the ISIC 2018 Task 3 image directory reaffirmed its prowess, showcasing a 70.41% accuracy. Despite the lower accuracy observed in testing against external images, I perceived this outcome as an opportunity for growth and optimization. It provided valuable insights into areas where enhancements and efficiencies can be made.
🌱 Future Directions: While the outcomes are promising, there's still room for refinement and expansion. Future iterations will focus on architectural enhancements, robust testing on diverse datasets, and integration of emerging technologies to further elevate diagnostic capabilities.
My inspiration for starting this project arises from my profound experience with machine learning projects and a desire to explore into the world of medical data. Seeking to push myself further, I began exploring and researching different paths, eventually deciding on skin cancer as the primary objective of my study. I am grateful to Pace University - Seidenberg School of Computer Science and Information Systems, IEEE Xplore, Stack Overflow, Medium.com, Kaggle.com, PapersWithCode and ChatGPT for providing the materials with insights that allowed me to fully understand the fundamental ideas of computer vision and transformer models.