<Background>

Breast cancer ranks as the most prevalent cancer among women globally. However, early detection contributes to a favorable prognosis for this form of cancer.  According to Cancer Information Service by National cancer center in Japan, While the incidence of cancer in women is the highest among other cancers, the 5-year survival rate is as high as 92.3% (2009-2011), and this underscores the favorable prognosis associated with early detection and treatment of the disease.  The continuum from identifying breast cancer to formulating a treatment plan initiates with clinical breast examination, followed by detection through imaging modalities such as mammography and ultrasound. Upon the identification of abnormalities, a biological classification is performed through cytological and histopathological analyses (Fig. 1).  In histopathology, the assessment of cancer presence or absence, its characteristics, and the extent of progression is ascertained from a tissue or cell sample extracted from a breast lesion identified through imaging studies at the specified location (Fig. 2).  Breast cancer subtypes are subsequently categorized based on the positivity or negativity of hormone receptors, specifically the estrogen receptor (ER) and progesterone receptor (PgR), as well as the human epidermal growth factor receptor 2 (HER2), a cell surface receptor (Fig. 3).  To date, cytological diagnosis has been employed for guiding treatment decisions; however, its drawback lies in its invasive nature, involving the insertion of a thick needle into the lesion and the excision of a tissue mass, imposing a substantial burden on the patient's body. 

Given the considerably invasive nature of histopathology, there has been ongoing discussion regarding the potential of vacuum-assisted biopsy (VAB) as a substitute for surgical specimens. Comparative studies have reported diagnostic concordance rates ranging from 79% to 100% for ER, 74% to 97% for PgR, and 60% to 98% for HER2, respectively (1, 2).  Hence, we hypothesized that classifying subtypes using mammograms taken for breast cancer detection could expedite the formulation of a treatment plan. Additionally, we envisaged prioritizing patients identified as triple-negative, characterized by the poorest prognosis and rapid progression, for early intervention. This approach aims to streamline examinations, facilitating the prompt initiation of treatment, which is called computer-aided triage (CADt).

In the domain of subtype classification through mammograms, Ma et al. have previously employed machine learning methodologies (3).  Their study utilized data from 331 Chinese women diagnosed with invasive breast cancer.  Naive Bayes machine learning model was performed with use of quantitative radiomics features such as morphology, grayscale statistics, and texture.  The study also revealed that the model integrating both CC and MLO images demonstrated greater accuracy compared to the individual use of each image type.

<Materials and Methods>

We randomly selected 84 cases from each of four subtypes (Luminal A, Luminal B, HER2-enriched, and triple-negative), matching cases of triple-negative, which has the lowest number, from The Chinese Mammography Database (CMMD).  A total of 336 selected cases were allocated into a 9:1 ratio for training (300 cases) and testing (36 cases) purposes. Cross-validation was performed on the training dataset to determine hyperparameters.

The computations were executed in Python v3.10.13 on a Linux server on Windows 11 Pro 64-bit system, 13th Gen Intel(R) Core(TM) i9-13900KF CPU at 3.00 GHz with 128 GB RAM, and NVIDIA GeForce RTX 4090 GPU 24 GB. The following Python packages were installed for the calculations: Torch v 2.0.1, LightGBM v4.1.0, Scikit-learn v 1.3.1.

In this investigation, we employed Vision Transformer (ViT), one of the latest deep learning models in image recognition, for feature extraction. Complementing this, Light Gradient Boosting Machine (LightGBM), known for its efficacy with small sample sizes, was integrated for classification, facilitating subtype classification based on mammograms.  Figure 4 illustrates the deep learning network model employed in this study.  Initially, two images captured in the MLO and CC directions were coalesced with rotation to align them at the center, and the resultant composite image was then subsampled to dimensions of 224x224.  Following this, the image underwent processing using pre-trained vision transformers (ViT) with vit_small_patch8_224.dino, embedding 384 features.  Subsequently, principal component analysis (PCA) was applied to compress the 384 features into 4 dimensions, defining them as ViT features.  In addition, PCA was used to obtain global features from the 224x224 image into 4 dimensions.  In conjunction with the ViT and global features, the age of patients and orientation (left or right) were incorporated to classify subtypes using LightGBM.  A comprehensive optimization of 10 LightGBM parameters was conducted utilizing Optuna (5), ver.3.3.0, across 1,000 iterations for each configuration. The parameter set yielding the highest F1 value was then employed for inference on the test data.