| Category | Specification |
|---|---|
| Primary Model | DenseNet121 (Baseline) |
| Hybrid Modules | Transformer Encoders, XGBoost |
| Environment | Kaggle (Dual Tesla T4 GPUs, 16GB VRAM) |
| Frameworks | TensorFlow 2.19.0, Keras, XGBoost |
| Primary Metric | Macro-F1 Score |
SCC-ALAR is a systematic study of convolutional and hybrid deep learning architectures for multi-class skin lesion classification using the HAM10000 dataset. The project emphasizes balanced evaluation, rigorous ablation studies, and clinically relevant performance under extreme class imbalance.
All experiments were conducted on the HAM10000 (Human Against Machine with 10000 training images) dataset, a standard benchmark for dermoscopic image analysis. The dataset contains 10,015 images across seven diagnostic categories.
- akiec – Actinic keratoses
- bcc – Basal cell carcinoma
- bkl – Benign keratosis-like lesions
- df – Dermatofibroma
- mel – Melanoma
- nv – Melanocytic nevi
- vasc – Vascular lesions
| Class | Image Count |
|---|---|
| nv | 6,705 |
| mel | 1,113 |
| bkl | 1,099 |
| bcc | 514 |
| akiec | 327 |
| vasc | 142 |
| df | 115 |
| Total | 10,015 |
- RGB conversion
- Resize to 224 × 224
- Pixel normalization to [0, 1]
Images were organized in a directory-based structure and loaded using TensorFlow’s image_dataset_from_directory.
A stratified split preserved class proportions:
- 70% Training
- 15% Validation
- 15% Test
| Class | Training | Validation | Test |
|---|---|---|---|
| nv | 4,693 | 1,006 | 1,006 |
| mel | 779 | 167 | 167 |
| bkl | 769 | 165 | 165 |
| bcc | 360 | 77 | 77 |
| akiec | 229 | 49 | 49 |
| vasc | 99 | 21 | 22 |
| df | 81 | 17 | 17 |
| Total | 7,010 | 1,502 | 1,503 |
To mitigate dominance of the nv class, inverse square-root class weighting was applied:
$ w_c = \frac{1}{\sqrt{f_c}} $
| Class | Weight |
|---|---|
| df | 9.303 |
| vasc | 8.415 |
| akiec | 5.533 |
| bcc | 4.413 |
| bkl | 3.019 |
| mel | 3.000 |
| nv | 1.222 |
-
DenseNet121 Backbone
Extracts a7 × 7 × 1024feature map. -
Tokenization
Feature map flattened into 49 tokens and projected to 256 dimensions. -
Transformer Encoder
- 2 layers
- 4 attention heads
- Models global spatial relationships.
-
Classifier Head
- MLP or XGBoost using extracted embeddings.
(Architecture diagram recommended here for visualization.)
- Accuracy: Overall correctness (biased toward majority class)
- Macro-F1 (Primary): Equal weighting across all classes
| Model | Test Accuracy | Test Macro-F1 |
|---|---|---|
| DenseNet121 | 0.7864 | 0.6029 |
| EfficientNetB0 | 0.6700 | 0.1242 |
| EfficientNetB3 | 0.6487 | 0.1476 |
| Custom CNN | 0.6660 | 0.3314 |
DenseNet121 clearly established itself as the strongest baseline.
| Class | Precision | Recall | F1-Score |
|---|---|---|---|
| nv | 0.90 | 0.90 | 0.90 |
| vasc | 0.83 | 0.68 | 0.75 |
| bcc | 0.59 | 0.73 | 0.62 |
| bkl | 0.60 | 0.62 | 0.61 |
| mel | 0.52 | 0.49 | 0.50 |
| df | 0.88 | 0.29 | 0.43 |
| akiec | 0.43 | 0.37 | 0.40 |
| Architecture | Test Accuracy | Test Macro-F1 |
|---|---|---|
| DenseNet121 (Baseline) | 0.7864 | 0.6029 |
| CNN → Transformer → XGBoost | 0.7731 | 0.5619 |
| CNN → Transformer → MLP | 0.7618 | 0.5427 |
| DenseNet → XGBoost | 0.7851 | 0.5131 |
Hybrid architectures did not outperform the CNN baseline, indicating overfitting or loss of spatial inductive bias.
├── assets/
├── densenet+trnsfrmr_embeddings/
├──densenet_embeddings/
├── experiment.ipynb
├── final_models_bundle/
│ ├── M1_DenseNet121_Classifier
│ ├── M2_DenseNet121_Backbone
│ ├── M3_CNN_Transformer_Encoder
│ └── M4_CNN_Transformer_MLP
└── README.md
This repository uses Git LFS to store model weights and embeddings.
git clone https://github.com/YOUR_USERNAME/SCC-ALAR.git
cd SCC-ALARpip install -r requirements.txtgit lfs pullfrom tensorflow import keras
import numpy as np
# Load the primary classifier
model = keras.models.load_model(
"final_models_bundle/M1_DenseNet121_Classifier/densenet121_classifier.keras"
)
# Prediction (expects a 224×224 RGB image normalized to [0, 1])
# predictions = model.predict(preprocessed_image)All experiments were conducted on the HAM10000 (Human Against Machine with 10000 training images) skin lesion dataset, a widely used benchmark for dermoscopic image classification. The dataset consists of 10,015 dermoscopic images categorized into seven diagnostic classes, representing both benign and malignant skin lesions.
The dataset includes the following classes:
- akiec – Actinic keratoses and intraepithelial carcinoma
- bcc – Basal cell carcinoma
- bkl – Benign keratosis-like lesions
- df – Dermatofibroma
- mel – Melanoma
- nv – Melanocytic nevi
- vasc – Vascular lesions
The class distribution is highly skewed toward benign lesions, particularly melanocytic nevi (nv), reflecting real-world clinical prevalence but posing a significant challenge for machine learning models.
Table 1: Overall Dataset Distribution
| Class | Label | Image Count |
|---|---|---|
| nv | Melanocytic nevi | 6,705 |
| mel | Melanoma | 1,113 |
| bkl | Benign keratosis-like lesions | 1,099 |
| bcc | Basal cell carcinoma | 514 |
| akiec | Actinic keratoses | 327 |
| vasc | Vascular lesions | 142 |
| df | Dermatofibroma | 115 |
| Total | 10,015 |
All images were:
- Converted to RGB format
- Resized to 224 × 224 pixels
- Normalized to the range [0, 1]
This resolution was chosen to ensure compatibility with standard ImageNet-pretrained CNN architectures while preserving sufficient lesion detail.
Images were organized into a directory-based class structure, enabling efficient loading via TensorFlow’s image_dataset_from_directory API.
To ensure fair evaluation and preserve class proportions, a stratified split was employed:
- 70% Training
- 15% Validation
- 15% Test
Table 2: Stratified Split Counts per Class
| Class | Training (70%) | Validation (15%) | Test (15%) |
|---|---|---|---|
| nv | 4,693 | 1,006 | 1,006 |
| mel | 779 | 167 | 167 |
| bkl | 769 | 165 | 165 |
| bcc | 360 | 77 | 77 |
| akiec | 229 | 49 | 49 |
| vasc | 99 | 21 | 22 |
| df | 81 | 17 | 17 |
| Total | 7,010 | 1,502 | 1,503 |
To improve generalization while avoiding excessive distortion of medically relevant features, light, conservative augmentation was applied only to the training set:
- Random horizontal flipping
- Small random rotations
- Mild zoom
- Limited contrast variation
Aggressive augmentations were avoided because dermoscopic color and texture are diagnostically meaningful, and excessive distortion can introduce non-physiological artifacts.
Given the dominance of the nv class, explicit imbalance handling was required to protect the recall of minority malignant classes.
Class weights were computed using an inverse square-root frequency scheme:
$ w_c = \frac{1}{\sqrt{f_c}} $
Table 3: Computed Class Weights for Training
| Class | Computed Weight |
|---|---|
| df | 9.303 |
| vasc | 8.415 |
| akiec | 5.533 |
| bcc | 4.413 |
| bkl | 3.019 |
| mel | 3.000 |
| nv | 1.222 |
- Accuracy: Measures overall correctness (biased toward majority classes).
- Macro-F1 Score: Primary metric; treats all classes equally to emphasize minority-class performance.
Table 4: CNN Benchmark Performance (Test Set)
| Model | Test Accuracy | Test Macro-F1 |
|---|---|---|
| DenseNet121 | 0.7864 | 0.6029 |
| EfficientNetB0 | 0.6700 | 0.1242 |
| EfficientNetB3 | 0.6487 | 0.1476 |
| Custom CNN | 0.6660 | 0.3314 |
Table 5: DenseNet121 Performance by Category
| Class | F1-Score | Recall (Sensitivity) |
|---|---|---|
| nv | 0.90 | 0.90 |
| vasc | 0.75 | 0.68 |
| bcc | 0.62 | 0.73 |
| bkl | 0.61 | 0.62 |
| mel | 0.50 | 0.49 |
| df | 0.43 | 0.29 |
| akiec | 0.40 | 0.37 |
Table 6: Final Architecture Comparison (Ablation Study)
| Model Architecture | Test Accuracy | Test Macro-F1 |
|---|---|---|
| DenseNet121 (Baseline) | 0.7864 | 0.6029 |
| CNN → Transformer → XGBoost | 0.7731 | 0.5619 |
| CNN → Transformer → MLP | 0.7618 | 0.5427 |
| DenseNet → XGBoost | 0.7851 | 0.5131 |
- HAM10000 exhibits extreme class imbalance.
- Conservative augmentation improves generalization without distorting features.
- DenseNet121 provides the best balance between accuracy and Macro-F1.
- Increased architectural complexity (Transformers/XGBoost) did not surpass the CNN baseline. These weights were applied during training across all deep learning models.
Due to class imbalance, Macro-F1 score was adopted as the primary evaluation metric, complemented by accuracy.
- Accuracy measures overall correctness but is biased toward majority classes.
- Macro-F1 treats all classes equally, emphasizing minority-class performance.
All reported metrics correspond to the held-out test set, ensuring unbiased evaluation.
Four CNN architectures were benchmarked:
- DenseNet121
- EfficientNetB0
- EfficientNetB3
- Custom CNN baseline
DenseNet121 achieved the highest test Macro-F1 (~0.60) and accuracy (~0.79), significantly outperforming EfficientNet variants, which exhibited majority-class collapse, and the custom CNN, which showed limited representational capacity. This established DenseNet121 as the strongest baseline for subsequent experiments.
Class-wise evaluation revealed:
- Strong performance on dominant and visually consistent classes (nv, bcc, bkl).
- Moderate performance on mel.
- Reduced recall and F1 scores for minority and visually ambiguous classes (akiec, df, vasc).
This behavior reflects both dataset imbalance and intrinsic diagnostic difficulty, reinforcing the importance of balanced evaluation.
Transformer-based hybrids and XGBoost classifiers were evaluated to test whether additional architectural complexity could improve performance. Despite perfect training scores in some cases, none of the hybrid approaches surpassed DenseNet121 in Macro-F1, indicating overfitting or loss of spatial inductive bias.
- HAM10000 exhibits extreme class imbalance, necessitating balanced metrics and weighting strategies.
- Conservative augmentation improves generalization without distorting medical features.
- Class weighting mitigates, but does not eliminate, minority-class performance gaps.
- DenseNet121 provides the best balance between accuracy and Macro-F1.
- Increased architectural complexity does not guarantee improved clinical performance.
These results motivate a deeper discussion on:
- The role of inductive bias in medical imaging.
- Limitations imposed by dataset scale and imbalance.
- Why Transformer-based global modeling may be ineffective in this context.
This study was designed as a controlled ablation analysis, where architectural complexity was incrementally increased to test whether it leads to improved balanced performance on an imbalanced medical imaging dataset.
- Ablation: Custom CNN vs EfficientNet (B0, B3) vs DenseNet121.
- Observation: DenseNet121 achieved the highest Macro-F1 (~0.60), while EfficientNet variants collapsed toward majority-class (nv) predictions, despite moderate accuracy.
- Interpretation: DenseNet’s dense connectivity promotes feature reuse and gradient flow, which appears particularly beneficial for small and visually subtle lesions. EfficientNet’s compound scaling struggles under severe class imbalance and limited dataset size.
- Key insight: Architectural efficiency does not necessarily translate to diagnostic robustness in imbalanced medical datasets.
- Ablation: DenseNet → CNN Tokenization → Transformer Encoder → MLP.
- Hypothesis: Transformer self-attention could model global spatial context and improve minority-class discrimination.
- Observation: Training performance improved, but validation and test Macro-F1 decreased. Minority-class recall did not improve.
- Interpretation: CNN feature maps already encode local spatial inductive bias. Flattening them into tokens may dilute spatial locality. The dataset scale is insufficient for learning meaningful long-range dependencies.
- Key insight: Global context modeling is not inherently beneficial when discriminative cues are localized and data is limited.
- Ablation: CNN → Transformer → Embeddings → XGBoost.
- Observation: Near-perfect training performance but no improvement in validation/test Macro-F1. Clear signs of overfitting.
- Interpretation: XGBoost excels at memorizing high-dimensional representations, but deep embeddings lack explicit structure for tabular learners.
- Key insight: Strong classifiers cannot compensate for representations that are not inherently separable for minority classes.
- Ablation: DenseNet → Global Embeddings → XGBoost.
- Observation: Performance comparable to CNN baseline with a slight reduction in Macro-F1.
- Interpretation: Replacing the neural head with XGBoost removes end-to-end optimization and class-weight–aware gradient updates.
- Key insight: End-to-end learning remains crucial for imbalanced medical classification.
- Observation: Several models achieved similar accuracy (~0.75–0.78), but Macro-F1 varied drastically (0.12–0.60).
- Interpretation: Accuracy was dominated by the nv class. Models with high accuracy but low Macro-F1 failed clinically relevant classes.
- Key insight: Accuracy is an insufficient metric for diagnostic systems; balanced metrics are mandatory.
Despite rigorous experimentation, this study has several limitations.
- 2.1 Dataset Scale and Imbalance: HAM10000 contains only ~10k images. Minority classes have fewer than 150 samples, limiting Transformer stability and minority-class generalization.
- 2.2 Absence of Lesion-Level Annotations: Only image-level labels were available. No lesion masks or region-of-interest annotations were used.
- 2.3 Single-Dataset Evaluation: All experiments were conducted on HAM10000 only. Cross-dataset generalization (e.g., ISIC archive) was not evaluated.
- 2.4 Limited Hyperparameter Exploration: Transformer depth and attention heads were intentionally constrained to maintain controlled ablations.
- 2.5 Lack of Clinical Metadata: No patient-level information (age, sex, lesion location) was used, which could significantly improve diagnostic performance.
The findings of this study suggest several promising directions for future research.
- 3.1 Data-Centric Improvements: Curated rebalancing using lesion-aware augmentation and synthetic minority oversampling.
- 3.2 Lesion-Focused Modeling: Integrating segmentation-based attention or multi-instance learning (MIL) using lesion patches.
- 3.3 Alternative Transformer Integration: Exploring hierarchical Transformers or local-window attention (e.g., Swin-style) trained jointly.
- 3.4 Cost-Sensitive and Recall-Oriented Loss Functions: Implementing Focal loss variants or class-specific recall penalties.
- 3.5 Cross-Dataset and Real-World Evaluation: Evaluating robustness to different acquisition devices, lighting variation, and skin tone diversity.
- 3.6 Explainability and Trust: Saliency and attention map validation to ensure clinical safety.
- In medical image classification, carefully designed CNNs with strong inductive bias and balanced evaluation outperform more complex hybrid architectures when data is limited and imbalanced.
- Negative results from Transformer-based ablations are scientifically meaningful, reinforcing the importance of data characteristics over architectural novelty.
This work presented a systematic investigation of convolutional and hybrid deep learning architectures for multi-class skin lesion classification using the HAM10000 dataset. Given the inherent class imbalance and clinical importance of minority malignant lesions, the study emphasized balanced evaluation metrics, particularly Macro-F1, over conventional accuracy.
A comprehensive benchmarking of convolutional neural networks demonstrated that DenseNet121 consistently outperformed EfficientNet variants and a custom CNN baseline. DenseNet’s dense connectivity facilitated effective feature reuse and stable gradient propagation, enabling superior minority-class discrimination under limited data conditions. In contrast, EfficientNet models exhibited majority-class collapse, highlighting the limitations of compound scaling strategies in highly imbalanced medical datasets.
To assess whether increased architectural complexity could improve performance, multiple ablation experiments were conducted using Transformer-based hybrids and feature-level XGBoost classifiers. Despite strong training performance, none of the hybrid models surpassed the DenseNet121 baseline on the test set. Transformer encoders applied to CNN feature maps failed to enhance balanced classification, while XGBoost classifiers overfit high-dimensional embeddings without improving generalization. These findings reinforce the conclusion that architectural novelty alone does not guarantee improved diagnostic performance, particularly when data is limited and discriminative cues are localized.
Class-wise analysis further revealed that while benign and common lesion categories were learned effectively, rare and visually ambiguous classes such as actinic keratoses and dermatofibroma remained challenging. This outcome reflects both dataset constraints and the intrinsic complexity of dermatological diagnosis, underscoring the need for cautious interpretation of automated systems in clinical settings.
Overall, this study highlights the importance of inductive bias, data-centric design, and appropriate evaluation metrics in medical image classification. The results demonstrate that a well-regularized CNN with balanced training strategies can outperform more complex hybrid architectures, providing a strong and interpretable baseline for future research. The insights gained from negative ablation results are equally valuable, guiding future efforts toward data quality, lesion-focused modeling, and clinically informed learning objectives rather than increased model complexity alone.