Convolutional Neural Network (CNN) in the domain of dermatology #

Author: Nikita Kurkulskiu
Full implementation on GitHub.

Introduction #

Skin-cancer screening with deep learning can reduce diagnostic workload, but black-box models hinder clinical adoption.
We fine-tune a ResNet-18 to separate melanoma from benign lesions on the open HAM10000 dermatoscopic dataset, then make its decisions explainable and auditable with T-CAV (Testing with Concept Activation Vectors). T-CAV quantifies how much human-defined concepts (age, sex, body location, diagnosis labels) drive predictions, exposing hidden bias and validating clinically relevant factors.

Background #

Classical explainability (Grad-CAM, LRP) highlights where a network looks, but does not answer what semantic attributes it uses.
T-CAV fills this gap:

1. Concept definition – A user provides a small image set exemplifying a concept (e.g., “male patients”).
2. CAV learning – A linear separator in latent space defines a Concept Activation Vector.
3. TCAV score – The directional derivative of the class logit along the CAV is computed; the proportion of positive derivatives estimates concept influence.
4. Bootstrap p-value – Random sign-flipped CAVs test statistical significance.

Section 1 – Dataset #

• HAM10000 (Kaggle): 10 015 dermatoscopic JPEGs across 7 diagnostic categories.
• We convert metadata into >100 concepts (age buckets, sex, localisation, diagnosis, etc.) using concept_prep.py.

Section 2 – Ethical Motivation #

Demographic or anatomical biases in datasets can translate into unfair or unsafe predictions. Quantifying concept influence lets us:

• Detect gender/age bias.
• Validate model focus on clinical features instead of artefacts (e.g., histology stamp).

Methodology #

3.1 Code Organisation #

Xai-Project/
├── archive/               # Raw HAM10000 images + metadata CSV
├── data_dir/              # train/val/test split (benign vs malignant)
├── concepts/              # >100 sub-folders with concept images
├── tcav_results/          # Auto-generated bar-charts & report
├── train.py               # CNN training script
├── data_prep.py           # Builds data_dir/
├── concept_prep.py        # Builds concepts/
├── tcav.py                # Basic TCAV (single class)
├── statistical_check.py   # Full pipeline: CAV, bootstrap, markdown report
└── TCAV_Report.md         # Auto-generated

3.1 Model & Code Snippets #

Training (excerpt from train.py)

model = models.resnet18(weights=models.ResNet18_Weights.IMAGENET1K_V1)
model.fc = nn.Linear(model.fc.in_features, 2)          # 2 classes
...
for epoch in range(epochs):
    train_loss, train_acc = train_one_epoch(...)
    val_loss, val_acc     = evaluate(...)

T-CAV scoring (excerpt)

def tcav_score(loader, model, capt, cav, cls_idx):
    cav_t = torch.from_numpy(cav).float().to(device)
    hits = total = 0
    with torch.enable_grad():
        for x,_ in loader:
            x = x.to(device).requires_grad_(True)
            model.zero_grad()
            model(x)[:, cls_idx].sum().backward()
            dd = (capt.act * cav_t).sum(1)             # directional derivative
            hits += (dd > 0).sum().item(); total += x.size(0)
    return hits / total

3.3 Pipeline #

flowchart LR
    A[data_prep.py] --> B(train.py)
    B -->|model_best.pth| C(statistical_check.py)
    C --> D[TCAV_Report.md<br>bar charts]
    E(concept_prep.py) --> C

Insights #

Key TCAV Findings (bootstrap p ≤ 0.05) #

No benign concepts passed p ≤ 0.05; highest benign concept was
localization_lower_extremity (score 0.984, p = 0.090).

Bias & Fairness Analysis #

• Gender: sex_male TCAV = 1 → network over-predicts melanoma for men.
• Age: Scores > 0.70 for age ≥ 60, but p > 0.14 – weak evidence.
• Location: Chest/trunk dominate; ear/hand minimal influence, indicating imbalance.

Conclusion #

• Achieved ~90 % validation accuracy with ResNet-18 on HAM10000.
• T-CAV uncovered pronounced demographic and anatomical biases beyond lesion visuals.
• Recommended next steps:
  1. Apply adversarial debiasing (prototype in statistical_check.py).
  2. Introduce visual ABCD concepts (asymmetry, border irregularity).
  3. Validate findings on external datasets (PH2, Derm7pt).