Authors Maxwell Chen and Kevin Miao

This repository contains our final project for CS194-80: Full Stack Deep Learning taught at UC Berkeley by Pieter Abbeel, Sergey Karayev and Joshua Tobin.

The deployment phase can be accessed here:

• Archive - Contains older versions and debugging notebooks/scripts.
• HAM10000_metadata.csv - Original metadata with diagnoses (unmodified from the original HAM10000 dataset)
• annotation-v2.py - This code contains the annotation script which outputs final.csv that is used as ground truth labels and bounding box areas by using the provided segmentation maps.
• annotation.py - Script for automatic annotation
• dataset.py - Pytorch dataset accompanied by transforms
• disc.ipynb/py - Debugging files
• final.csv - Ground Truth bounding box coordinates, paths and labels
• mean-std.pt - PyTorch Dictionary containing the mean/std of the training images
• model_util.py - Util functions for loading/reading models from state
• setup.sh - Shell script for setting up requirements and dependencies
• state_loading.py - Script for loading a state dictionary into a model
• sweep.yaml - Weights and Biases files for hyperparameter sweeping
• train.ipynb - Notebook for training debugging
• train.py - Official training script
• transforms.py - Image Transforms
• util.py - Contains util functions

Tschandl, Philipp, 2018, "The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions", https://doi.org/10.7910/DVN/DBW86T, Harvard Dataverse, V3, UNF:6:/APKSsDGVDhwPBWzsStU5A== [fileUNF]

The basis of the project is the HAM10000 dataset which contains 10,015 images categorized in 7 different skin cancers along with supervised segmentation maps.

Dataset can be downloaded here: Harvard Dataverse

The architecture being used is a pretrained FasterRCNN with a ResNet50 backbone augmented with a Feature Pyramid Network. The model has been adapted from: torchvision FasterRCNN Resnet-50 fpn pretrained on COCO

We have the following diseases in our dataset which correspond to the respective indices. The last index, 7, is reserved for background.

Dictionary : {'akiec': 0, 'bcc': 1, 'bkl': 2, 'df': 3, 'mel': 4, 'nv': 5, 'vasc': 6}

This part of the project uses python 3.8 in a conda environment with the following dependencies. The setup.sh file can be run to initiate the online environment.