Melanoma classification from dermoscopic images

1. Clinical question and intended use

Dermoscopy magnifies the pigment structures of a skin lesion, and melanoma is the skin cancer that dermoscopic triage most needs to catch. This example asks a narrow research question: how well can a small convolutional network, trained from scratch on 540 labeled dermoscopic images, rank new images toward the melanoma label?

The example shows the complete Neural Designer workflow for image classification: importing an image folder, training a network, choosing an operating threshold on validation data and reporting every error on testing data. It does not establish clinical performance.

2. Images, labels and endpoint

The data are the 900 dermoscopic images of the ISIC 2016 Challenge, Part 3 training release, published by the International Skin Imaging Collaboration under CC0 1.0. Each image has one label: 727 are benign and 173 are malignant (melanoma), a melanoma share of 19.2%. The endpoint is that label; the positive class is malignant.

The original JPEG images range from 576 × 542 to 4288 × 2848 pixels (median 1024 × 768). Each one is converted to RGB and resized to 300 × 300 pixels with Lanczos resampling, without cropping, so non-square lesions are slightly distorted. This is the same preparation as the 102-image OpenNN example, whose images are bit-identical to the corresponding files here.

Neural Designer imports the folder with one subfolder per class and assigns images at random to training (60%), validation (20%) and testing (20%) subsets. The project stores this assignment.

3. Model

The network is the default image-classification architecture that Neural Designer 8.0.1 creates for 300 × 300 RGB images: one convolutional layer with eight 3 × 3 filters, max pooling and a single sigmoid output. It has 180,225 trainable parameters.

Output contract. The output is a score between 0 and 1 for the malignant class; the benign score is its complement. The scores are not calibrated probabilities. The image is flagged as suspicious when the malignant score reaches the decision threshold described in Section 5.

4. Training strategy

Training minimizes the cross-entropy error with the adaptive moment estimation (Adam) optimizer: learning rate 0.0001, mini-batches of 32 images, L2 regularization with weight 0.01 and 100 epochs. Training ran on an NVIDIA GeForce RTX 5070 Ti GPU.

The training and validation errors below are the mean binary cross-entropy of the malignant score over the images, without the regularization term. For reference, a constant score of 0.5 gives 0.693.

The training error falls from 0.753 to 0.261, while the validation error stays between about 0.40 and 0.50 after the first epochs. The network keeps fitting the training images without improving on unseen ones: with 540 training images, the data limit what this architecture can learn. The saved model is the one at epoch 100.

5. Model selection and threshold

Five configurations were trained on the same split and compared by their final validation error. The testing images were not used for this choice.

The first four candidates differ by at most 0.031, which is within the epoch-to-epoch fluctuation of the validation error (up to 0.08), so the choice among them is weak; the selected model keeps the regularization. The larger network with a hidden layer of 128 neurons did not train: its error stayed near 3 and its malignant score saturated at 0 for every image checked.

Decision threshold

Melanomas are 19% of the training images, so the model’s scores are low: the median malignant score is 0.075 for benign and 0.171 for malignant testing images. The default threshold of 0.5 therefore flags very few images. The threshold was set on the validation images, as the point of the validation ROC curve closest to perfect classification: 0.11, with validation ROC AUC 0.750, sensitivity 76.5% and specificity 65.8%.

Baseline. Predicting “benign” for every image is correct for 143 of the 180 testing images (79.4% accuracy) but detects no melanoma. Accuracy alone is therefore not a useful measure here.

6. Clinical validation

The final model is evaluated once on the 180 testing images: 37 malignant (20.6%) and 143 benign. Neural Designer reports a testing ROC AUC of 0.739. Calculated from the same scores, the 95% confidence interval is about 0.64–0.84 (Hanley–McNeil).

Operating point at threshold 0.11

At 0.5 the model behaves almost like the all-benign baseline: same accuracy, and only 4 of 37 melanomas detected. At 0.11 it detects 26 of 37 melanomas, at the cost of 45 false alarms among 143 benign lesions. On the testing images, the ROC point closest to perfect classification lies at a threshold of 0.13. That value is descriptive only, because it was found on the testing data.

7. Workflow and reproducibility

A responsible real-world analogue would begin after a lesion has entered an approved clinical pathway. It would check image quality and format, calculate a versioned score, and route flagged and unflagged lesions alike to clinician review, with histopathology as the reference where indicated.

Calculating outputs for new images

Neural Designer’s Calculate outputs task scores one image at a time. The two testing images below show one melanoma above the threshold and one benign lesion below it.

Example 1

Malignant score: 0.8151
Flagged: the score is above the threshold of 0.11.

Class	Output score
benign	0.1849
malignant	0.8151

Example 2

Malignant score: 0.0209
Not flagged: the score is below the threshold of 0.11.

Class	Output score
benign	0.9791
malignant	0.0209

These are illustrative outputs for images already counted in the testing metrics. The scores are uncalibrated.

Reproduce the example

The ZIP contains the 900 prepared images, the manifest, the trained Neural Designer 8.0.1 model (model/melanoma.ndm and its parameter file), the Neural Designer inference script and scores.csv with the malignant score of every image and its subset. All the metrics above can be recalculated from scores.csv. To score a new image prepared in the same way, run:

model\melanoma_predict.bat lesion.bmp --table

The script reports its predicted class at 0.5; apply the 0.11 threshold to the malignant score instead. To retrain, create an image-classification project in Neural Designer from the melanoma_dataset_bmp folder, keep the default network and use the training settings in Section 4. Neural Designer draws a new random split and initialization, so the results will vary; the saved split is listed in model/melanoma.ndm and scores.csv.

8. Safety, generalizability and governance

• Internal random split only. Training, validation and testing images come from the same release. No external hospital, device, time period or prospective cohort is evaluated.
• Patient grouping cannot be verified. The release has no patient identifiers, so related lesions may appear in more than one subset, which can make internal results optimistic.
• Few melanomas in the test. With 37 testing melanomas, each missed case changes sensitivity by 2.7 percentage points, and the confidence interval of the AUC is wide.
• Challenge selection, not a clinical population. The 19% melanoma share reflects how the challenge images were collected. Predictive values change with prevalence.
• Image preparation. Downsampling to 300 × 300 discards fine dermoscopic structure, the resize distorts non-square images, and no color or artifact normalization is applied.
• Uncalibrated scores. A threshold of 0.11 shows that the scores are not probabilities of melanoma.
• No subgroup evidence. Skin type, age, sex, anatomical site and acquisition device are not available in the release, so performance across them is unknown.
• No clinical-utility evaluation. The example does not compare with dermatologists or measure the effect on patient outcomes.

References

• ISIC 2016 Challenge, Part 3 training release. International Skin Imaging Collaboration (ISIC). CC0 1.0 Universal.
• Gutman D, Codella NCF, Celebi E, Helba B, Marchetti M, Mishra N, Halpern A. Skin Lesion Analysis toward Melanoma Detection: A Challenge at the International Symposium on Biomedical Imaging (ISBI) 2016, hosted by the International Skin Imaging Collaboration (ISIC). arXiv:1605.01397, 2016.
• Prepared images, trained Neural Designer 8.0.1 model and scores, 6 October 2026 (ZIP).
• More Neural Designer examples