TT-Stack: A Transformer-Based Tiered-Stacking Ensemble Framework with Meta-Learning for Automated Breast Cancer Detection in Mammography

1 st Showkat Osman, 2 nd Md. Tajwar Munim Turzo, 3 rd Maher Ali Rusho *, 4 th Md. Makid Haider, 5 th Sazzadul Islam Sajin, 6 th Ayatullah Hasnat Behesti, 7 th Ahmed Faizul Haque Dhrubo, 8 th Md. Khurshid Jahan, 9 th Mohammad Abdul Qayum · 2025

Read TT-Stack: A Transformer-Based Tiered-Stacking Ensemble Framework with Meta-Learning for Automated Breast Cancer Detection in Mammography by 1 st Showkat Osman, 2 nd Md. Tajwar Munim Turzo, 3 rd Maher Ali Rusho *, 4 th Md. Makid Haider, 5 th Sazzadul Islam Sajin, 6 th Ayatullah Hasnat Behesti, 7 th Ahmed Faizul Haque Dhrubo, 8 th Md. Khurshid Jahan, 9 th Mohammad Abdul Qayum (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-02091.txt.

Introduction

Breast cancer remains a leading cause of cancer death, and catching it early in mammograms saves lives. Showkat Osman and eight co-authors present TT-Stack in 2025, a transformer-based ensemble framework that combines several models with a meta-learner for automated detection in mammography images.

Older computer-aided diagnosis systems struggled to represent rich features and to generalize across varied mammographic images. Stacking diverse models in tiers and letting a meta-model weigh their outputs aims at both problems. The framework's goal is dependable automated screening support, where a missed cancer and a false alarm both carry real costs.

Catalog record

CREATOR1 st Showkat Osman, 2 nd Md. Tajwar Munim Turzo, 3 rd Maher Ali Rusho *, 4 th Md. Makid Haider, 5 th Sazzadul Islam Sajin, 6 th Ayatullah Hasnat Behesti, 7 th Ahmed Faizul Haque Dhrubo, 8 th Md. Khurshid Jahan, 9 th Mohammad Abdul Qayum
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-02091

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TT-Stack: A Transformer-Based Tiered-Stacking Ensemble Framework with Meta-Learning for Automated Breast Cancer Detection in Mammography 1 st Showkat Osman, 2 nd Md. Tajwar Munim Turzo, 3 rd Maher Ali Rusho *, 4 th Md. Makid Haider, 5 th Sazzadul Islam Sajin, 6 th Ayatullah Hasnat Behesti, 7 th Ahmed Faizul Haque Dhrubo, 8 th Md. Khurshid Jahan, 9 th Mohammad Abdul Qayum Abstract Breast cancer continues to be the second most common cause of cancer-related deaths around the world, with early detection being important to improve survival rates for patients. Traditional computer-aided diagnosis systems have limitations in their ability to represent features and generalize to the range of mammographic images. We present a new two-level Stack of Transformers (TT-Stack) ensemble framework based on using heterogeneous lightweight vision transformer architectures to automatically identify breast cancer in mammograms. Specifically, we integrate seven state-of-the-art vision transformers: RepViT, DaViT, EfficientViT, MobileViT, FasterViT, MViT, and PVT-v2 while also designing a two-tier meta-learning approach for the ensemble by simply taking the logits from the base model and applying logistic regression for binary classification (Cancer vs. Non-Cancer). Each of the transformer backbone models was developed to process single-channel grayscale mammograms while still taking advantage of transfer learning from pre-training on ImageNet so that they would offer a parameter-efficient approach that may reasonably be applied in clinical practice with minimal variance. The training process included stratified 80/20 splits when necessary, class-balanced upsampling, early stopping, and an adaptive learning rate schedule on the public Mammogram Mastery dataset. In separate evaluations here, it was determined that EfficientViT and PVT-v2 were the top performing models achieving 99.33% validation, 97.96% F1-score, and perfect 1.000:0 ROC-AUC with only small train/validation gaps ( ≤ \leq 0.67 points). Finally, the TT-Stack ensemble model by the end of the evaluation reached 99.33% accuracy with 100% precision, 96% recall, 97.96% F1-score and a 99.97% ROC-AUC, and demonstrated robustness in performance due to the diversity of the architecture. In comparison with the latest state-of-the-art methods, our framework demonstrates improved performance; for example, TT-Stack achieved above the performance of RadiNet-XGBoost (96.67%) and MedFoundX (98.7%) on the same dataset. The framework also establishes a novel benchmark for transformer-based breast cancer detection while being computationally efficient through lightweight architectures of the method, creating a clear pathway to creating scalable, accurate mammography screening systems, particularly in resource-limited clinical contexts. I Introduction Bangladesh, with a population of approximately 173.97 million as of 2024 (Worldometer), faces significant health challenges, with nearly half of the population comprising women [ 1 ] . Each year, a large number of people suffer from various diseases, with women disproportionately affected due to societal norms that inhibit them from openly discussing health issues. In many cases, women refrain from sharing their medical concerns, even with family members or healthcare providers, fearing stigmatization and being perceived as a burden. As a result, numerous women succumb to preventable conditions due to delayed diagnosis and treatment. Breast cancer, in particular, stands as one of the most critical health issues faced by women, both globally and in Bangladesh. Breast cancer is the second leading cause of cancer-related death worldwide. It occurs when breast cells undergo mutations, becoming cancerous and multiplying to form tumours. Although the disease predominantly affects women, especially those assigned female at birth (AFAB) and aged 50 and older, it can also impact men and younger individuals [ 2 ] . Approximately 80% of breast cancer cases are invasive, meaning the tumour may spread beyond the breast to other parts of the body [ 3 ] . Breast cancer is classified into common and less common types. Common types include invasive ductal carcinoma, lobular breast cancer, and ductal carcinoma in situ. Less common types include triple-negative breast cancer, inflammatory breast cancer, and Paget’s disease of the breast [ 3 ] . Recognizing the early signs and symptoms of breast cancer is crucial for effective treatment. Common symptoms include changes in breast size, shape, or contour, persistent lumps or thickening in the breast or underarm, and discharge from the nipple, which may be blood-stained or clear. Early detection significantly improves treatment outcomes. Various imaging techniques are available for screening and diagnosing breast cancer, with mammography, ultrasound, and thermography being the most prevalent. Mammography, which uses X-rays, is the gold standard for early detection, while magnetic resonance imaging (MRI) is employed for high-risk individuals. Physical examinations, such as clinical breast exams performed by healthcare professionals and self-examinations, also play a vital role in early detection. Women’s awareness of changes in their breasts, such as alterations in size, lumps, or pain, is essential in prompting timely medical consultation. In this work, we present a Transformer-based Tiered-Stacking (TT-Stack) ensemble framework for mammogram-based breast cancer classification that will systematically combine seven lightweight vision transformer architectures (RepViT, DaViT, EfficientViT, MobileViT, FasterViT, MViT, and PVT v2) that each have properties for grayscale medical imaging. These proposed models have several important contributions with regard to: • Proposing a unique two-tier meta-learning ensemble that takes each of the base-model logits concatenated into a 14-dimensional feature vector, then refines that with logistic regression, while taking advantage of differences in architectures (i.e. depthwise separable convolutions, dual attention, hierarchical multi-scale features for features) for better discriminative power. • Introducing efficient grayscale-optimized transformer adaptation (in_chans=1) which can transfer learning from ROS-ready pre-trained weights, while retaining better parameter efficiency for clinically-ready use. • Employing a stratified class-balanced training protocol which utilized upsampling-based balancing on the training set only to prevent leakage (and included a validation and test set), while implementing early stopping and ReduceLROnPlateau scheduling for eta. • Benchmarking all models (engagement with transparency) in a multi-model framework (ROC-AUC, precision/recall, confusion-matrix) to establish performance baselines for transformer-based breast cancer detection. • Establishing a modular pipeline ready for production (preprocessing, augmentation, reproducible training (SEED=42), and ensemble). In contrast to the dominant methods for single-architecture models, the importance of TT-Stack frameworks for reducing a gap in medical AI is that lightweight transformers can be integrated into an ensemble, ultimately demonstrating that there can be improved generalization from fractional and complementary feature learning, while retaining computational fidelity to be practicable for world-ready screening workflows. II Literature Review The detection of breast cancer has been thoroughly researched with various imaging modalities and machine learning techniques with the goal of improving the accuracy of diagnostics. In this section, we consolidate the most relevant investigations by organizing our sections around methodologies, imaging strategies, and comparisons of the advantages and disadvantages of studies. Sadoughi et al. [ 4 ] , developed an extensive framework for detecting breast cancer that incorporated mammography, ultrasound, and thermography. They reported high accuracy of diagnosis using support vector machines (SVM) through the analysis of over 40 datasets. They rigorously investigated texture features, edge-based descriptors, and fuzzy techniques to compare the benefits and limitations of imaging modalities. In another study on mammographic microcalcifications, Prannoy and Saravanakumar [ 5 ] , developed a computer-aided diagnosis (CAD) framework for diagnostic triage of the lesions between benign, normal and malignant. They use Principal Component Analysis (PCA) to reduce the dimensionality of the data they collected from DDSM database with approximately 3,000 samples, suggesting CAD use can be used to detect tumors responsible for the mammographic microcalcifications in cancer screening workflows. Neural

Read the complete text