Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lightweight Deep Models for Dermatological Disease Detection: A Study on Instance Selection and Channel Optimization

Ian Mateos Gonzalez, Estefani Jaramilla Nava, Abraham Sánchez Morales, Jesús García-Ramírez, Ricardo Ramos-Aguilar

TL;DR

This paper tackles dermatological disease detection from RGB skin images using the DermaMNIST dataset and aims to achieve competitive performance with lightweight CNNs suitable for CPU-only training. It introduces a two-phase approach: dataset analysis with instance selection and channel optimization, followed by CNN evaluation under computational constraints. The study shows that instance selection improves class separation, but retaining all three RGB channels yields the best accuracy among lightweight models, achieving 71.57% on DermaMNIST—close to the 73.5% reported for larger ResNet models—while drastically reducing parameters to ~472K. These results demonstrate the practicality of low-resource, GPU-free dermatology screening and highlight directions for further improving class balance and model robustness.

Abstract

The identification of dermatological disease is an important problem in Mexico according with different studies. Several works in literature use the datasets of different repositories without applying a study of the data behavior, especially in medical images domain. In this work, we propose a methodology to preprocess dermaMNIST dataset in order to improve its quality for the classification stage, where we use lightweight convolutional neural networks. In our results, we reduce the number of instances for the neural network training obtaining a similar performance of models as ResNet.

Lightweight Deep Models for Dermatological Disease Detection: A Study on Instance Selection and Channel Optimization

TL;DR

This paper tackles dermatological disease detection from RGB skin images using the DermaMNIST dataset and aims to achieve competitive performance with lightweight CNNs suitable for CPU-only training. It introduces a two-phase approach: dataset analysis with instance selection and channel optimization, followed by CNN evaluation under computational constraints. The study shows that instance selection improves class separation, but retaining all three RGB channels yields the best accuracy among lightweight models, achieving 71.57% on DermaMNIST—close to the 73.5% reported for larger ResNet models—while drastically reducing parameters to ~472K. These results demonstrate the practicality of low-resource, GPU-free dermatology screening and highlight directions for further improving class balance and model robustness.

Abstract

The identification of dermatological disease is an important problem in Mexico according with different studies. Several works in literature use the datasets of different repositories without applying a study of the data behavior, especially in medical images domain. In this work, we propose a methodology to preprocess dermaMNIST dataset in order to improve its quality for the classification stage, where we use lightweight convolutional neural networks. In our results, we reduce the number of instances for the neural network training obtaining a similar performance of models as ResNet.

Paper Structure

This paper contains 10 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Proposed Methodology
  3. Exploratory image analysis
  4. Evaluation
  5. DermaMNIST Analysis
  6. Instances for each class
  7. Data visualization
  8. Channel selection
  9. Experimental Results
  10. Conclusions and Future Work

Figures (4)

  • Figure 1: Distribution of the classes in DermaMNIST dataset. The acronyms corresponds to the diseases in the dataset: Melanocytic Nevi (MN), Melanoma (ME), Benign Keratosis-like lesions (BK), Basal Cell Carcinoma (BC), Actinic Keratoses (AL), Vascular Lesions (VL), Dermatofibroma (DF)
  • Figure 2: Distribution of the classes in DermaMNIST dataset. The acronyms corresponds to the diseases in the dataset: Melanocytic Nevi (MN), Melanoma (ME), Benign Keratosis-like lesions (BK), Basal Cell Carcinoma (BC), Actinic Keratoses (AL), Vascular Lesions (VL), Dermatofibroma (DF)
  • Figure 3: Visualization of the data with t-SNE and Isomap algorithms. (a) Isomap before the transformation; (b) Isomap after the transformation; (c) t-SNE before the transformation; (d) t-SNE after the transformation.
  • Figure 4: Visualization of the correlation among the RGB channels. (a) Scatter plot displaying the mean values of the RGB channels, where each point represents an instance in the dataset; (b) Correlation matrix of the three channels. A strong correlation between the green and blue channels is observed.