PaperlayerView exact revision on arXiv

Source-linked AI summary

Early Prediction of Pathological Complete Response to Neoadjuvant Chemotherapy Using Temporal Deep Learning on DWI

Pablo García Marcos, Md. Tarequl Islam, Paula Puerta González, Guillermo Lorenzo, Héctor Gómez, Covadonga del Camino, Angel Rio-Alvarez, Víctor M. González

arXiv:2609.26106v1cs.AI

TL;DR

This paper addresses the limited availability of early pCR prediction approaches using diffusion imaging alone. It applies temporal deep learning to tumor-centered imaging from baseline and after the first NACT cycle, achieving an AUC of 0.90 ± 0.06 across 10-fold cross-validation. The authors identify cohort size and reliance on expert tumor masks as limitations.

  • Problem

    Many response-prediction models focus on later treatment stages or use multimodal inputs, while early prediction with standalone diffusion imaging remains methodologically underexplored.

    Source: §I · p. 1, §II · p. 3

  • Method

    The study applies temporal deep learning to tumor-centered DW-MRI data from baseline and after the first NACT cycle.

    Source: §I · p. 2, §II · p. 3

  • Results

    AUC was 0.90 ± 0.06 across 10 folds for the EfficientNet-B0 and bidirectional LSTM model using tumor-centered patches.

    Source: §V · p. 7

  • Takeaways & Limitations

    The results indicate the feasibility of early NACT-response prediction using information from before treatment and during the first cycle.

    Source: §V · p. 7

    The 169-patient cohort may limit statistical robustness, and expert-annotated tumor masks reduce clinical scalability.

    Source: §V · p. 7

Abstract

from arXiv · show

Early identification of non-responders to neoadjuvant chemotherapy (NACT) is crucial for timely treatment adaptation in breast cancer. However, many existing predictive models rely on multiparametric magnetic resonance imaging (MRI), late treatment time points, or extensive clinical data, which limits their applicability. This study proposes a deep learning framework for early prediction of pathological complete response (pCR) using only diffusion-weighted MRI (DW-MRI) acquired at baseline and after the first NACT cycle. This framework feeds cropped tumor-centered patches to an EfficientNet-based temporal model that directly learns tumor shape and local tissue characteristics without explicit radiomic feature engineering. The model, trained with 10-fold cross-validation, achieved an area under the receiver operating characteristic curve (AUC) of 0.90 for pCR prediction after one cycle, providing actionable information after a single treatment cycle while avoiding gadolinium administration and reducing dependence on heterogeneous clinical data. By focusing on the baseline-to-first-cycle window instead of later stages, the approach supports earlier escalation or de-escalation of NACT, and its exclusive reliance on DW-MRI facilitates protocol standardization, multi-centre deployment and privacy-preserving data sharing. These results demonstrate that DW-MRI-based deep learning on tumor-centered patches constitutes a minimally invasive, clinically deployable strategy for early pCR prediction, with direct implications for personalized treatment adaptation in neoadjuvant breast cancer therapy.

I. INTRODUCTION

Early pCR prediction matters because many patients do not achieve pCR, while predictions made late in NACT may come too late to guide meaningful treatment changes. This study develops a model using DW-MRI from baseline and after the first NACT cycle.

I. INTRODUCTION

  • Achieving pCR means no residual disease in the breast and lymph nodes and is strongly associated with improved survival in HER2-positive and triple-negative tumors.NACT can reduce tumor size, facilitate surgery, indicate treatment sensitivity, and treat micrometastases early.Source: §I · p. 1
  • Late-stage response evaluation may leave too little time for meaningful therapy changes, motivating prediction during early treatment.The introduction links early response estimates to treatment tailoring and reduced unnecessary toxicity.Source: §I · pp. 1–2
  • DCE-MRI requires gadolinium and adds cost, scan time, personnel and material needs, and safety concerns for some patients.Differences in imaging between centers and heterogeneous multimodal inputs also complicate standardization and reproducibility.Source: §I · p. 2
  • DW-MRI and ADC provide information about tissue properties, and longitudinal ADC changes have been associated with pathologic response during NACT.The text presents standalone DW-MRI or derived ADC maps as a potential simpler alternative to contrast-enhanced protocols.Source: §I · p. 2
  • The study targets early prediction of pCR at surgery from DW-MRI acquired at baseline and after the first NACT cycle.It proposes temporal deep learning on patches of tumor-affected tissue, using DW-MRI and derived ADC maps.Source: §I · p. 2

II. RELATED WORK

Prior pCR prediction work spans radiomics, deep learning, and multimodal models, with trade-offs in timing, contrast use, data requirements, and deployment complexity. The paper addresses a gap in early prediction using diffusion imaging alone.

II. RELATED WORK

  • Radiomics-ML pipelines can use DCE or diffusion descriptors, but often require manual segmentation and heterogeneous clinical inputs, with scanner and protocol variability complicating generalization.These practical constraints motivate simpler approaches aimed at reproducibility and multicenter generalization.Source: §II · pp. 2–3
  • DCE-MRI deep-learning models have often used pretreatment or post-NACT images, but late treatment time points limit their usefulness for adaptation.Gadolinium constraints and sensitivity to acquisition domain shifts also motivate contrast-free alternatives.Source: §II · p. 3
  • Multimodal models combine imaging with clinical, histological, or biomarker data and may improve AUC and sensitivity, while increasing pipeline complexity and data requirements.Inter-center acquisition differences and heterogeneous data availability can hinder deployment and increase vulnerability to domain shifts.Source: §II · p. 3
  • Although diffusion imaging has shown potential for response prediction, most deep-learning approaches combine it with DCE-MRI or other inputs rather than use DW-MRI alone.The review describes a methodological gap in early, diffusion-only deep learning using data gathered from multiple centers.Source: §II · p. 3
  • This work uses DWI at baseline and after one NACT cycle, learning from tumor-centered patches to pursue early prediction with lower logistical load and easier standardization.The paper positions the approach against pipelines relying on contrast and multiple sequences.Source: §II · p. 3

III. MATERIALS AND METHODS

The study uses the ACRIN 6698 multicenter trial dataset, which includes MRI examinations across treatment time points and expert tumor segmentations for diffusion images and ADC maps.

A. Data Description

  • ACRIN 6698 includes 385 patients with up to four MRI examinations: before treatment, after three weeks of paclitaxel, between paclitaxel and anthracycline, and after four anthracycline cycles.The dataset was released as part of a multicenter clinical trial focused on NACT for breast cancer.Source: A. Data Description · p. 3
  • Each MRI session includes DCE-MRI, DW-MRI, ADC maps, and expert manual tumor segmentations for DW-MRI and ADC.The segmentations are binary masks, with tumor as the positive class and other tissues as negative.Source: A. Data Description · pp. 3–4
  • Fig. 1 presents one study slice as DW-MRI, ADC map, and mask, respectively, from left to right.The mask identifies the tumor region corresponding to the imaging slice.Source: Figure 1 · p. 4, A. Data Description · p. 4

B. Pre-Processing

The preprocessing pipeline pairs images with masks, normalizes and crops volumes, and prepares ADC-map inputs for a temporal model that extracts features and models changes across time points.

B. Pre-Processing

  • The preprocessing pipeline pairs each ADC and DW slice with its mask, resizes slices to 256 × 256, normalizes volumes, and crops them to 3 × 50 × 50 patches.Patients without all diffusion studies are removed, and non-tumorous ADC tissue is masked out with a five-pixel margin around tumor borders.Source: B. Pre-Processing · p. 4
  • Fig. 2’s arrows encode patients × time points × slices × width × height, allowing dataset dimensions to be compared across preprocessing steps.The described transformations include resizing, volume reconstruction, and cropping.Source: Figure 2 · p. 4, B. Pre-Processing · p. 4

C. Methods

  • The model uses EfficientNet-B0 features, a bidirectional LSTM for longitudinal changes, and temporal attention to weight time points before classification.The classifier also uses image quality indices; the stated input images are ADC maps.Source: 1) Deep Learning Architectures · p. 4, C. Methods · p. 4

1) Deep Learning Architectures:

The architecture uses a tumor-centered imaging model to predict NACT response, with training and evaluation details specified for the final model.

1) Deep Learning Architectures:

  • The final model’s architecture is shown as the NACT response prediction model.The model produces a binary classification using two linear layers, ReLU activation, and 40% dropout.Source: 1) Deep Learning Architectures · p. 5, Figure 3 · p. 5
  • Training used 169 patients, with T0 and T1 3×50×50 images, and oversampling to address class imbalance.Training ran up to 100 epochs with early stopping; augmentation included flips, rotations, translations, scaling, and mild intensity changes.Source: 2) Training, Validation and Test Data · p. 5

2) Training, Validation and Test Data:

  • The final model was reported using 10-fold cross-validation, with 90% of patients used for training in each fold.Development ablations instead used 5-fold cross-validation and approximately 80% of patients for training.Source: 2) Training, Validation and Test Data · p. 5
  • Performance was evaluated against ground-truth annotations using accuracy, F1-score, and AUC-ROC.AUC-ROC measures discriminative ability across decision thresholds.Source: 3) Performance Metrics · p. 5

3) Performance Metrics:

Performance measures include accuracy, while the supplied table caption identifies a backbone and temporal-module ablation evaluated using mean validation accuracy and F1-score.

3) Performance Metrics:

  • Table I compares backbones and temporal modules on 50 × 50 tumor patches using mean validation accuracy and F1-score across folds.The ablation uses the T0–T1 setting, 5-fold cross-validation, and no quality indices.Source: Table I · p. 5
  • Accuracy is the proportion of correct predictions among all true-positive, true-negative, false-positive, and false-negative outcomes.The equation defines accuracy as (TP + TN) divided by (TP + TN + FP + FN).Source: 3) Performance Metrics · p. 5
  • For binary NACT-response prediction, pCR is designated as the positive class.TP, TN, FP, and FN denote true positive, true negative, false positive, and false negative, respectively.Source: 3) Performance Metrics · p. 5

IV. RESULTS

The ablation compares recurrent and non-recurrent temporal approaches and identifies an EfficientNet-B0 LSTM configuration as the strongest reported backbone combination.

1) Temporal Modeling and Backbone:

A. Ablation Study

  • Across backbones, LSTM temporal modeling matched or outperformed non-recurrent aggregation, supporting explicit temporal dynamics with only two time points.The non-recurrent approach concatenated features from T0, T1, and their difference.Source: 1) Temporal Modeling and Backbone · p. 5

1) Temporal Modeling and Backbone:

  • Table III evaluates adding image-quality indices to imaging features from tumor-centered patches in the T0–T1 EfficientNet-B0 LSTM setting.The caption specifies 5-fold cross-validation; no outcome values are provided in the supplied passages.Source: Table III · p. 6
  • 76% accuracy and 0.77 F1 were achieved by EfficientNet-B0 with LSTM, versus 75% accuracy and 0.76 F1 for its simple temporal version.These configurations were evaluated in the T0–T1 backbone ablation using 5-fold cross-validation on 50 × 50 tumor patches without quality indices.Source: 1) Temporal Modeling and Backbone · pp. 5–6
  • A second ablation explored LSTM hidden-state sizes and dropout rates with the selected EfficientNet-B0 LSTM configuration held fixed.Its goal was to maximize validation F1 while maintaining stability across folds.Source: 2) Temporal Module Hyperparameters · p. 6

2) Temporal Module Hyperparameters:

The selected LSTM settings were chosen based on validation performance, while image-quality indices were designed to augment the temporal image representation.

2) Temporal Module Hyperparameters:

  • 64 hidden units and 0.40 dropout yielded the highest mean validation F1-score without signs of overfitting, and were adopted for later experiments.Increasing hidden size or dropout slightly improved performance across tested LSTM configurations.Source: 2) Temporal Module Hyperparameters · p. 6
  • Image-quality indices summarize MRI and segmentation reliability, including signal-to-noise properties, motion artifacts, and tumor-region temporal consistency.The lesion-level indices are concatenated with features from the backbone and temporal module to modulate predictions according to acquired-data confidence.Source: 3) Quality Indices · p. 6

3) Quality Indices:

The final T0–T1 model achieved competitive early pCR prediction performance, while ablations found modest gains from quality indices and favored EfficientNet-B0 with bidirectional LSTM temporal modeling. Validation performance varied across folds, and ViT-based models scored lowest among the tested architectures.

3) Quality Indices:

  • Adding quality indices increased mean accuracy by approximately one percentage point while leaving F1-score essentially unchanged; the limited 5-fold ablation was not conclusive.The authors retained the indices for their clinical interpretability and reevaluated them in the final configuration.Source: 3) Quality Indices · p. 6

B. Experimental Results and Analysis

  • The final model showed high inter-fold validation variability, which the authors associate with the small validation subsets and influence of individual patient outcomes.Table IV reports per-fold validation performance for the final T0–T1 model with quality indices.Source: B. Experimental Results and Analysis · p. 6, Table IV · p. 6
  • ViT-based backbones yielded the lowest scores among tested architectures with both recurrent and non-recurrent temporal modules.The authors relate this pattern to the larger datasets and stronger augmentation typically needed for vision transformers, compared with convolutional backbones in small-sample medical imaging.Source: B. Experimental Results and Analysis · p. 7
  • 84.3% (SD ≈8.18) accuracy and 0.85 (SD ≈0.07) F1-score with quality indices, versus 82.24% (SD ≈8.33) accuracy and 0.83 (SD ≈0.08) F1-score without them.With quality indices, performance matched or improved on the majority of folds in 10-fold cross-validation.Source: B. Experimental Results and Analysis · p. 7
  • AUC = 0.90 in cross-validation, exceeding ADC-based approaches and slightly improving on some multiparametric methods, but remaining below.The authors note that comparisons are difficult because studies use different inputs and techniques.Source: B. Experimental Results and Analysis · p. 7

V. CONCLUSIONS AND FUTURE WORK

The study reports feasible early NACT-response prediction from information available before treatment and during the first cycle, using tumor-centered patches and a temporal model. The authors identify a small cohort and reliance on expert tumor masks as limitations, and propose further validation and model extensions.

V. CONCLUSIONS AND FUTURE WORK

  • Across 10 cross-validation folds, the EfficientNet-B0 plus bidirectional LSTM model achieved accuracy of 84.30 ± 7.08, F1-score of 0.85 ± 0.07, and AUC of 0.90 ± 0.06.The model used tumor-centered patches and information from before treatment and during the first cycle.Source: §V · p. 7
  • The authors report moderate fold variability, with most folds’ AUC typically above ≈0.88, and describe the multicenter dataset as suggesting potential for generalization.They recommend further evaluation in multicenter clinical settings.Source: §V · p. 7
  • Future work could test predictions using baseline-only inputs, add clinical variables or imaging sequences, and externally validate on independent centers.The authors say any performance gains from additional data should be weighed against the disadvantages of heterogeneous inputs.Source: §V · p. 7
  • The final cohort of 169 patients may limit statistical robustness, and expert-annotated tumor masks reduce clinical scalability.The authors identify automated segmentation as a needed direction.Source: §V · p. 7

ETHICS STATEMENT

The study used a publicly available, fully de-identified imaging dataset for secondary analysis. The authors report that IRB approval and informed consent were not required under the dataset’s use policies.

ETHICS STATEMENT

  • The study exclusively used the ACRIN 6698 dataset, a publicly available, fully de-identified breast cancer imaging dataset from TCIA.The dataset is identified by accession number tcia.kk02-6d95.Source: ETHICS STATEMENT · pp. 7–8
  • Under TCIA data use policies, IRB approval and informed consent were not required for this secondary analysis of de-identified, publicly available data.Source: ETHICS STATEMENT · p. 8
  • The authors state that no ethical complications were identified during the study.Source: ETHICS STATEMENT · p. 8
Loading 2609.26106v1…