💻 computer science

A Preliminary CNN Baseline for Breast Ultrasound Classification in MATLAB, with Exploratory IDC/ILC Labels: Toward Explainable Breast Imaging AI

This paper establishes a preliminary MATLAB-based CNN baseline for binary malignant versus non-malignant breast ultrasound classification using the BrEaST dataset, achieving improved test accuracy while explicitly framing exploratory subtype labeling and explainability as future research directions rather than validated results.

Original authors: ISHANI CHOVATIYA

Published 2026-07-28

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: ISHANI CHOVATIYA

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the human body as a vast, mysterious city, and doctors as the detectives trying to solve its most critical mysteries. Sometimes, the clues are hidden deep inside, where a regular flashlight can't reach. That's where medical imaging comes in, acting like a super-powered X-ray vision that lets doctors peek inside without making a single cut. Among these tools, breast ultrasound is like a friendly, non-invasive scout that uses sound waves to create pictures of what's happening inside the breast tissue. It's great at spotting lumps, but telling the difference between a harmless bump (like a bubble in a cake) and a dangerous one (like a storm cloud) can be tricky. The images are often fuzzy, depend on who is holding the probe, and look different every time. This is where Artificial Intelligence (AI) steps in, hoping to act as a tireless, super-observant sidekick that can spot patterns humans might miss. But for AI to be a trusted partner in a hospital, it needs to be trained carefully, tested rigorously, and—most importantly—explain why it makes a decision, rather than just guessing.

This paper is like a "Day One" report from a young researcher named Ishani Chovatiya, who is building a very first version of that AI sidekick. Think of it as constructing the skeleton of a robot before adding the muscles or the brain. The researcher used a public collection of breast ultrasound images called the BrEaST dataset, which contains 256 scans. She taught a type of AI called a Convolutional Neural Network (CNN)—which is basically a digital brain designed to look at pictures—to sort these images into two piles: "Malignant" (dangerous) and "Non-Malignant" (safe). She did all this using MATLAB, a popular software tool for math and engineering.

The results are a mix of "good start" and "still learning." When the AI first tried to guess, it got about 70.88% of the answers right. After some tweaking and training, the improved version got up to 78.43% accuracy. That's a step forward, like a student improving their test score, but the paper is very clear: this is not ready for a real hospital yet. In fact, the AI got worse at spotting the dangerous cases (a metric called "recall") while getting slightly better at being precise. It's like a security guard who stops fewer innocent people but also misses a few actual threats. The researcher also tried to see if the AI could tell the difference between two specific types of cancer (IDC and ILC) just by reading the text notes in the dataset, but there were so few examples that she couldn't give a reliable score for that part.

Perhaps the most honest part of the paper is what it doesn't do. The researcher admits that this AI is a "black box"—it makes a guess, but we don't know exactly which part of the image made it think "danger." She didn't use any special tools to make the AI explain its reasoning, and she explicitly states that this is a preliminary baseline, not a finished product. The goal isn't to claim victory, but to lay down a clear, documented starting point. It's a map showing where the journey began, with a promise that future work will add the missing pieces: better training, a fixed plan so the results can be repeated exactly, and the ability for the AI to say, "I think this is cancer because I see these specific shapes," so doctors can trust it. Until then, this AI is a promising student, not yet a certified doctor.

Technical Summary: A Preliminary CNN Baseline for Breast Ultrasound Classification in MATLAB

Problem Statement
Breast ultrasound is a critical modality for assessing breast lesions, yet automated classification remains challenging due to image noise, operator dependency, and visual variability. While recent research has advanced toward multimodal report generation using Transformers, there is a need for documented, reproducible baselines that establish foundational workflows for simpler classification tasks. This paper addresses the lack of a transparent, conservative baseline for binary breast lesion classification (malignant vs. non-malignant) using the BrEaST dataset, while explicitly acknowledging the limitations of current explainability and subtype classification efforts.

Methodology
The study utilizes the Breast-Lesions-USG (BrEaST) dataset from The Cancer Imaging Archive, comprising 256 patient-level scans (154 benign, 98 malignant, 4 normal). The workflow, implemented in MATLAB, involves the following steps:

Label Preparation: A binary classification scheme was created where "malignant" cases are positive, and "benign" and "normal" cases are grouped as non-malignant. An exploratory component attempted to identify Invasive Ductal Carcinoma (IDC) and Invasive Lobular Carcinoma (ILC) subtypes via text matching in the "Diagnosis" field, identifying 14 cases for each.
Data Preprocessing: Images were resized to 224 x 224 pixels, normalized using training-set statistics, and augmented via random reflection, translation, rotation, and scaling to mitigate overfitting.
Model Architecture: A Convolutional Neural Network (CNN) was constructed using the MATLAB Deep Learning Toolbox. The architecture includes convolutional layers, batch normalization, ReLU activation, max-pooling, dropout, fully connected layers, and a softmax output.
Training Configuration: The model was trained for 50 epochs using the Adam optimizer (initial learning rate $1 \times 10^{-4}$ ) with a mini-batch size of 32. An 80/20 train-test split was used, though the specific split indices were not preserved with a fixed random seed.
Explainability: No explainability method (e.g., Grad-CAM, LIME) was applied to the final reported CNN model. An exploratory LIME-style decision-tree analysis exists in supplementary code but is explicitly excluded from the paper's validated results.

Key Results
The paper reports a conservative final comparative result to avoid presenting contradictory development-stage metrics:

Performance Improvement: The improved model achieved a test accuracy of 78.43%, up from an earlier model's 70.88%.
Precision: Increased from 0.600 to 0.625.
Recall and F1-Score: These metrics decreased in the final model (Recall: 0.385; F1-score: 0.476), indicating that while overall accuracy improved, the model's ability to detect positive cases and balance precision/recall requires further tuning.
IDC/ILC Analysis: No quantitative accuracy was reported for the IDC/ILC exploratory labels. Due to the small sample size (14 cases each) and the lack of a preserved fixed-seed test split for these specific subtypes, the authors deemed the results unreliable for publication as a validated classifier.

Key Contributions

MATLAB CNN Baseline: A documented workflow for binary breast ultrasound classification using the BrEaST dataset, covering data loading, label creation, augmentation, and training.
Conservative Reporting: A focus on a single, final comparative test accuracy (78.43%) rather than competing development-stage metrics, prioritizing transparency over inflated performance claims.
Exploratory Subtype Identification: An initial attempt to derive IDC and ILC labels from unstructured diagnosis text, highlighting the potential for future subtype-focused research while clarifying current limitations.
Roadmap for Explainability: A clear delineation of the current work as a "first historical baseline," explicitly positioning explainable AI (XAI) and uncertainty estimation as necessary future extensions rather than current achievements.

Significance and Claims
The paper positions itself not as a clinically ready diagnostic system, but as a methodological and developmental baseline. The authors emphasize that the 78.43% accuracy is insufficient for clinical deployment and that the drop in recall and F1-score underscores the risk of relying solely on accuracy in medical imaging.

The primary significance lies in providing a transparent, reproducible starting point for future research. The work serves as a foundation for:

Implementing fixed-seed reruns to ensure reproducibility.
Integrating visual explainability methods (e.g., Grad-CAM).
Developing uncertainty-aware models.
Advancing toward multimodal image-text modeling for report generation.

The study concludes that while the current CNN baseline demonstrates the ability to learn patterns from the BrEaST dataset, it requires external validation, stronger architectures, and clinical review before it can contribute to trustworthy breast imaging AI.

Technical Summary: A Preliminary CNN Baseline for Breast Ultrasound Classification in MATLAB

More like this