AI for Mammography: Can 512-Pixel Patches Preserve Detail?

High-resolution mammograms preserve fine detail, but processing an entire image can demand substantial computing resources. This study examines whether smaller image patches can make CNN-based classification more practical, comparing two neural-network architectures on digital mammograms. The validation results are promising, but they also show why patch-level classification is not the same as a clinically validated assessment of a full examination.

Key findings

• The study compared ResNeXt50 and EfficientNet-b6 for binary classification of mammography patches.
• Each patch measured 512 × 512 pixels and came from the VinDr-Mammo dataset.
• Both models achieved 0.78 accuracy and 0.78 recall on the validation data.
• EfficientNet-b6 achieved 0.83 precision, compared with 0.78 for ResNeXt50.
• The paper reports a more stable validation-loss curve for EfficientNet-b6, while combining patch predictions at patient level remains an open challenge.

Why high-resolution mammography is difficult to classify

Mammography is an important method for breast cancer screening and early detection. Full-field digital mammography, or FFDM, provides high-resolution images, but those images can be difficult to interpret. Overlapping tissue, differences in breast density and subtle appearances of abnormalities all add complexity. Some cases can be challenging even for experienced readers.

Convolutional neural networks, or CNNs, learn patterns in images and can be trained to classify them. In mammography, a central engineering problem is how to preserve small details without making computation impractical. Resizing a complete image can remove information from small affected regions. Feeding the full-resolution image into a sufficiently large model, however, requires considerable memory and processing capacity. The paper investigates image patches as a way to manage this trade-off.

Dataset and patch preparation

The study uses VinDr-Mammo, a dataset containing 20,000 DICOM images from 5,000 studies. The dataset includes finding information such as masses, calcifications, asymmetry and other disorders. Findings are marked with bounding boxes, and the data also include breast density and BI-RADS rankings. The paper states that the images were anonymized and that legal and ethical principles were followed during data collection.

For classification, finding categories were grouped into one positive class, while images with no finding formed the negative class. The study focused on the most prevalent MONOCHROME 2 photometric interpretation because differences between mammography machine image interpretations caused training difficulties. The paper reports 4,500 training samples, with 900, or 20%, used for validation.

A patch is a cropped part of an image that is processed as an individual model input. The selected patch size was 512 × 512 pixels. For positive examples, the centre of a finding’s bounding box was used as a reference for cropping. Images without findings had no bounding boxes, so the image centre of mass was calculated and random offsets were added when selecting negative patches. The images were normalized, and training augmentation included random horizontal or vertical flips and rotations from 0 to 180 degrees.

The paper also reports tests of reducing the size of complete images, but that approach was abandoned after poor preliminary results. The paper attributes this in part to the loss of information from small affected areas during resizing. Patch extraction reduced memory demands and made it possible to train larger and deeper models than whole-image processing would allow on the available hardware.

Models, training and validation results

ResNeXt and EfficientNet served as CNN feature-extraction backbones. In simple terms, the backbone learns image features, which are then passed to a classification head to predict whether a patch belongs to the positive or negative class. The models were trained for 20 epochs using PyTorch and the Adam optimizer. Batch sizes were 128 for ResNeXt and 16 for EfficientNet. Training took place on a workstation equipped with a 48 GB NVIDIA RTX 6000 Ada Generation GPU.

On the validation set, ResNeXt50 achieved 0.78 accuracy, 0.78 precision, 0.78 recall and an F1 score of 0.78. Accuracy measures the share of correct predictions overall. Precision indicates how many predicted positive cases were correct, while recall measures how many actual positive cases were detected. F1 combines precision and recall into one metric.

EfficientNet-b6 achieved 0.78 accuracy, 0.83 precision, 0.78 recall and an F1 score of 0.78. In other words, the two models had the same reported accuracy, recall and F1 score, while EfficientNet-b6 had higher precision. The paper also describes EfficientNet-b6’s validation-loss curve as more stable and continuously decreasing, in contrast with the more variable ResNeXt50 curve. The paper suggests that attention layers in EfficientNet might help it classify very small abnormalities. This is the paper’s proposed explanation for the results, not a separately established finding from the comparison.

What the results mean for practice

For engineers and researchers, the study illustrates a practical trade-off between image detail and computational demand. Patches of 512 × 512 pixels retain local image information while requiring less memory than processing entire high-resolution mammograms. This approach may support the development of AI methods for medical image analysis, but the reported results alone do not establish that a system is ready for clinical use.

For students, it is important to distinguish a patch classification result from an assessment of a complete mammography examination. The models were trained to classify patches as positive or negative. The paper identifies the integration of patch predictions into a patient-level result as an unresolved task. Although bounding boxes helped define positive crop locations, the reported metrics do not separately measure how accurately the model localizes findings.

Limitations and next steps

The results table reports validation metrics but does not provide a separate test-set evaluation or external validation on another dataset. The scores therefore do not demonstrate how the models would perform across different institutions or clinical conditions. In addition, different finding types were merged into one positive class, so the models were not evaluated on their ability to distinguish among those categories. The paper notes that accuracy needs further improvement before clinical implementation.

One proposed next step is to develop an additional model that combines patch-level predictions into a patient-level assessment. Further research would also need to assess that approach on other data and examine how reliably it identifies abnormalities. Such work would address a key gap between classifying cropped image regions and supporting the interpretation of a complete examination.

Conclusion

This study shows how 512 × 512-pixel patches can help train CNN classifiers on high-resolution mammograms while reducing computational demands compared with processing full images. EfficientNet-b6 achieved higher precision than ResNeXt50, while their accuracy, recall and F1 scores were the same. The comparison offers a useful research result for AI-based mammography classification, while highlighting the need to combine patch predictions and conduct further evaluation before clinical use.

Frequently asked questions

Author: Adi Pandžić, PhD in Mechanical Engineering

This article summarises a published research paper; please refer to and cite the original paper. About the author

← All articles