proposal.org (3105B)
1 #+TITLE: Diabetic Retinopathy Classification 2 #+AUTHOR: Vineet Kumar 3 #+DATE: 4 #+OPTIONS: H:2 num:t toc:t \n:nil @:t ::t |:t ^:t -:t f:t *:t <:t 5 #+OPTIONS: TeX:t LaTeX:t skip:nil d:nil todo:t pri:nil tags:not-in-toc 6 #+INFOJS_OPT: view:nil toc:nil ltoc:t mouse:underline buttons:0 path:https://orgmode.org/org-info.js 7 #+EXPORT_SELECT_TAGS: export 8 #+EXPORT_EXCLUDE_TAGS: noexport 9 #+HTML_LINK_UP: 10 #+HTML_LINK_HOME: 11 #+startup: beamer 12 #+LaTeX_CLASS: beamer 13 #+LaTeX_CLASS_OPTIONS: [aspectratio=169,bigger] 14 #+COLUMNS: %40ITEM %10BEAMER_env(Env) %9BEAMER_envargs(Env Args) %4BEAMER_col(Col) %10BEAMER_extra(Extra) 15 #+BEAMER_THEME: metropolis 16 #+latex_header: \usepackage{fontspec} 17 #+latex_header: \setsansfont{Comic Neue} 18 19 * Objective 20 ** Objective 21 - Classifying diabetic retinopathy into 5 classes 22 1. Healthy retina 23 2. Mild nonproliferative diabetic retinopathy 24 3. Moderate nonproliferative diabetic retinopathy 25 4. Severe nonproliferative diabetic retinopathy 26 5. Proliferative diabetic retinopathy 27 - Input would be a picture of a retina (like those from an optometrist 28 when they scan your eye) 29 30 * Materials and Method 31 ** Materials 32 - Using a dataset from Kaggle ("Diabetic Retinopathy Arranged") that 33 has all the labelled images: 34 - There is a class imbalance (73% / 25k images are of healthy 35 retinas), would be accounted for with the loss function when 36 training 37 - Class 0: 25k images 38 - Class 1: 2.4k images 39 - Class 2: 5.2k images 40 - Class 3: 800 images 41 - Class 4: 700 images 42 43 ** Image Examples 44 *** figure 45 :PROPERTIES: 46 :BEAMER_col: 0.5 47 :END: 48 #+ATTR_LATEX: :width 0.8\textheight 49 #+CAPTION: Class 0: Healthy retina 50 [[./retina_class0.jpg]] 51 52 *** figure 2 53 :PROPERTIES: 54 :BEAMER_col: 0.5 55 :END: 56 #+ATTR_LATEX: :width 0.8\textheight 57 #+CAPTION: Class 4: Proliferative diabetic retinopathy 58 [[./retina_class4.jpg]] 59 60 61 ** Method 62 - Training a CNN model for classification 63 - There exists a few that can be finetuned on a dataset such as 64 Microsoft's ResNet-50 (2015) and Google's EfficientNet-B0 (2019) 65 - I would be comparing with both models 66 - Preprocessing 67 - Resize images to 224x224 (input size that EfficientNet uses) 68 - Apply CLAHE (Contrast Limited Adaptive Histogram Equalization) on 69 luminosity channel of LAB 70 - there is more usable information for a retina in the luminosity 71 channel than the chrominance (colour) channel 72 - I would also compare without CLAHE to see how that affects 73 metrics 74 - Normalize with ImageNet's mean and std 75 - Perform random flips and rotations of the images so that the model 76 can generalize across orientations 77 78 ** Metrics 79 *** Primary 80 - Weighted F1-score (to account for class imbalance) 81 *** Additional 82 - Per-class precision and recall 83 - Confusion matrix 84 85 * Timeline 86 ** Timeline 87 - Day 1-3: dataset setup, make preprocessing pipeline, perform baseline 88 training on both models 89 - Day 4-6: evaluate and compare outputs, tune weighted loss function, 90 maybe train with additional epochs for better metrics (depending on 91 how long it takes) 92 - Day 7-9: gather final metrics, start preparing final report and 93 presentation slides