AI xPress

AI and computer science research.

Back to Foragentis

Healthcare & Bio AI

A deep learning model classified breast cancer in histopathology images with 93.97 percent accuracy and in ultrasound images with 89.87 percent accuracy

Healthcare & Bio AI November 6, 2026

A retina camera in an empty clinic room

In a 2025 paper in Scientific Reports, researchers describe a model called Deep Neural Breast Cancer Detection, or DNBCD. It classified histopathology images, which are microscope images of tissue, with 93.97 percent accuracy. It classified breast ultrasound images with 89.87 percent accuracy. The model also offers visual explanations for its predictions. Researchers on the paper list affiliations at several universities, including Bangladesh Army University of Science and Technology.

How accurate was the model?

The researchers evaluated the model on two image collections. The first, Breakhis-400x, also called B-400x, holds 1,820 histopathology images and has benign and malignant classes. The second, Breast Ultrasound Images Dataset, or BUSI, holds 1,578 ultrasound images and has benign, malignant, and normal classes. They split each collection into 70 percent for training, 20 percent for testing, and 10 percent for validation. The model reached 93.97 percent accuracy on B-400x and 89.87 percent accuracy on BUSI.

How did the researchers make the model explainable?

The system uses a deep network called Densenet121 as its base and adds customized convolutional layers. The researchers used image normalization, resizing, and augmentation, which creates altered copies of training images. They also used class weights to address class imbalance. The model uses Grad-CAM to show which parts of an image influenced its prediction. The authors say these visual justifications can make the model more transparent to healthcare providers.

What are common questions about this study?

Did the study test the model in a clinic?

No. The abstract reports evaluation on two benchmark image collections, not a clinical test with patients.

Why do visual explanations matter?

They let a clinician see what the model looked at when it made a decision. The authors say that can increase trust and transparency.

What did the authors make public?

The authors state that they made all the materials publicly accessible for the research community.

Limits and what's next

The evaluation used two existing image collections: 1,820 histopathology images in B-400x and 1,578 ultrasound images in BUSI. The abstract reports no independent replication or outside validation on new patient images. The design was a computational benchmark with training, testing, and validation splits, not a clinical study. The authors describe potential uses in clinical environments, but the abstract does not report testing in those environments. They also state that the materials are publicly accessible.

Source: Md. Romzan Alom et al., “An explainable AI-driven deep neural network for accurate breast cancer detection from histopathological and ultrasound images,” Scientific Reports, 2025.