Frequently Asked Questions

DeepMammo currently accepts PNG and JPG/JPEG images up to 20 MB. DICOM files (.dcm) should be converted to PNG using your imaging software before uploading.

DeepMammo was trained on 45,304 augmented images from the CBIS-DDSM and INbreast datasets. The model uses EfficientNet-B4 with FPN and CBAM attention. Exact metrics (dice score, IoU, classification accuracy) are available in the project documentation. This is a research demo — it has not been clinically validated.

Abnormality type tells you what kind of finding is present: a mass (a solid lump) or calcification (calcium deposits). Pathology predicts whether the finding is likely benign (non-cancerous) or malignant (potentially cancerous). Both come with confidence scores.

The segmentation head of the model predicts a pixel-level binary mask highlighting the abnormal region in the mammogram. The overlay blends this mask onto the original image in red so you can see exactly which area the model is flagging.

The saliency map shows where the model is paying attention. A jet colormap (blue → red) is used — red areas have high model attention, blue areas have low. It is derived from the segmentation probability map, providing a spatial explanation of the model's decision.

The clinical report is generated by a large language model (Llama 3 70B via Groq) using the model's structured predictions as input. It produces a radiologist-style narrative with sections for clinical indication, findings, impression, and recommendation. It is AI-generated and must be reviewed by a qualified radiologist.

No. DeepMammo is a research and demonstration tool. It has not been approved for clinical use, has not undergone regulatory review, and should never be used as the sole basis for a medical decision. Always consult a qualified radiologist or medical professional.

No. Images are processed in memory and not saved to disk or any database. Once the response is returned, the image data is discarded.