https://github.com/gtkacz/transform-ordering-selection

Code for my undergraduate thesis: Quantitative Analysis of the Impact of Image Pre-Processing on the Accuracy of Computer Vision Models Trained to Identify Dermatological Skin Diseases
https://github.com/gtkacz/transform-ordering-selection

cnn computer-vision machine-learning pre-processing preprocessing

Last synced: 24 days ago
JSON representation

Code for my undergraduate thesis: Quantitative Analysis of the Impact of Image Pre-Processing on the Accuracy of Computer Vision Models Trained to Identify Dermatological Skin Diseases

• Host: GitHub
• URL: https://github.com/gtkacz/transform-ordering-selection
• Owner: gtkacz
• License: mit
• Created: 2024-03-28T00:18:24.000Z (about 2 years ago)
• Default Branch: main
• Last Pushed: 2026-04-13T03:04:13.000Z (about 1 month ago)
• Last Synced: 2026-04-13T05:08:36.268Z (about 1 month ago)
• Topics: cnn, computer-vision, machine-learning, pre-processing, preprocessing
• Language: Jupyter Notebook
• Homepage: https://www.overleaf.com/read/kzmvsyrpncgh#ca0f54
• Size: 3.01 GB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          

    


        


        
Significance of Transform Ordering and Selection in CNN Preprocessing for Binary Skin Classification

        
Gabriel Mitelman Tkacz, Gustavo Scalabrini Sampaio, Leandro Augusto da Silva

        
Mackenzie Presbyterian University — São Paulo, Brazil

    




## Abstract

Image preprocessing is a near-universal step in medical image classification pipelines, yet its impact on model accuracy remains insufficiently characterized, particularly with respect to the ordering of multi-transform compositions. This study reports an exhaustive combinatorial evaluation of all 65 pipelines constructible from four fundamental transforms — histogram equalization, intensity normalization, non-local means denoising, and color space conversion — applied to a four-block convolutional neural network trained on the binary healthy-versus-diseased dermatoscopic classification task formed by using a dataset with 20,000 balanced images. Every pipeline was trained from scratch under five independent random seeds. Against a near-ceiling baseline of 98.09% test accuracy, which arithmetically bounds the detectable positive-α regime to at most +1.91 pp, no pipeline of length 2 or greater achieves a positive mean accuracy gain, and only equalization alone produces a seed-robust improvement. Holm-corrected permutation tests establish three ordering regularities: pipeline length correlates negatively with accuracy gain (*r* = −0.44), equalization-first placement outperforms alternative first-position assignments by 1.10 pp (Cohen's *d* = 1.22), and the bookend configuration with equalization first and normalization last outperforms its mirror image by 2.00 pp (Cohen's *d* = 2.38). At pipeline *k* = 3, ordering accounts for 58% of accuracy-gain variance and transform selection for 42%. Confusion-matrix decomposition indicates that preprocessing-induced degradation is dominated by bidirectional feature-space damage (78%) rather than recoverable decision-threshold shift (22%). Within the scope of this single network, single binary task, single dataset, and single transform set, the evidence supports a cautious harm-minimization heuristic — equalization first, normalization last — rather than a general claim about ordering dominance in medical image preprocessing.

## Keywords

Image preprocessing · Transform ordering · Preprocessing pipeline design · Convolutional neural networks · Medical image classification · Dermatological diagnosis