PaperlayerView exact revision on arXiv

Source-linked AI summary

Neoadjuvant chemotherapy response prediction using pretreatment diffusion and contrast-enhanced magnetic resonance imaging with clinical variables

Pablo García Marcos, Paula Puerta González, Guillermo Lorenzo, Héctor Gómez, Covadonga del Camino, Adán Rodríguez, Ignacio Peláez, Angel Rio-Alvarez, Víctor M. González

arXiv:2609.26105v1cs.AIcs.CV

TL;DR

Pretreatment pCR prediction could inform planning before NACT, but it must infer tumor sensitivity without treatment-induced changes. This study combines baseline ADC, DCE-MRI, and clinical information in a multimodal model; the combined configuration achieved an AUC of 0.86 in cross-validation. The authors conclude that external validation is needed before clinical implementation.

  • Problem

    Baseline pCR prediction could guide initial treatment decisions, but is challenging because it lacks treatment-induced dynamics and reliable predictors remain limited.

    Source: §1 · pp. 1–2

  • Method

    The study proposes a pretreatment pCR model combining ADC maps, DCE-MRI, and clinical variables.

    Source: §1 · p. 2

  • Results

    ADC+DCE-MRI+HR/HER2 subtype achieved the best AUC, reaching 0.86 in cross-validation.

    Source: §6 · p. 5

  • Takeaways & Limitations

    Pretreatment multimodal learning can estimate response before NACT begins, without waiting for treatment-induced changes.

    Source: §5 · p. 4

    External validation is needed to validate model performance for clinical implementation.

    Source: §6 · pp. 5–6

Abstract

from arXiv · show

Prediction of pathological complete response before neoadjuvant chemotherapy may facilitate more tailored therapeutic planning for breast cancer patients. This work proposes a deep-learning model for pretreatment data only, combining apparent diffusion coefficient maps, dynamic contrast-enhanced magnetic resonance imaging, and clinical variables. The study uses the public ACRIN 6698/I-SPY2 multicenter dataset. The architecture employs EfficientNet-B0 pretrained encoders for image feature extraction and late fusion with clinical information. Multiple clinical variables were evaluated, including age, race, histological type, HR/HER2 subtype, SBR grade, and maximum diameter. Only HR/HER2 subtype improved the average area under the receiver operating characteristic curve (AUC) and was retained in the final model. Using stratified five-fold cross-validation, standalone apparent diffusion coefficient maps achieved a mean AUC of 0.79, whereas dynamic contrast-enhanced magnetic resonance imaging achieved 0.74. Adding HR/HER2 subtype improved performance to 0.83 and 0.81, respectively. The final configuration, using both imaging modalities and HR/HER2 subtype, achieved an AUC of 0.86. These results support pretreatment multimodal learning for response prediction, although external validation is required before clinical use.

1. Introduction

The introduction frames pCR as an important NACT outcome but notes that reliable response predictors are lacking. It motivates pretreatment MRI and clinical data as a route to prediction before treatment begins.

1. Introduction

  • NACT can reduce tumor volume and assess treatment sensitivity, while pCR is associated with better prognosis, especially in HER2-positive and triple-negative tumors.Without reliable predictors, some patients may receive NACT toxicity without proportional benefit.Source: §1 · p. 1
  • Baseline pCR prediction could guide initial treatment intensity or adaptive-therapy selection, but is more challenging because it lacks treatment-induced dynamics.Unlike longitudinal prediction, it can be made before the first treatment cycle.Source: §1 · p. 2
  • DCE-MRI captures vascular perfusion and permeability, whereas diffusion-derived ADC maps characterize water mobility and tumor cellularity.The introduction presents these imaging sources as complementary inputs that may improve prediction when combined with clinical variables.Source: §1 · p. 2
  • The study proposes a T0-only pCR model combining ADC maps, DCE-MRI, and clinical variables, and evaluates each modality and the clinical variables' impact on performance.Its stated contributions include a standardized pretreatment input pipeline and a multimodal neural network.Source: §1 · p. 2

2. Related work

Prior pCR prediction work includes radiomics and deep learning, but often relies on standardized imaging or information acquired during or after NACT. The section motivates a baseline approach combining ADC, DCE-MRI, and limited clinical data.

2. Related work

  • Radiomics extracts tumor texture, shape, and intensity features for supervised classification, but depends on stable acquisition protocols and inter-scanner harmonization.Integrating clinical and molecular data can improve discrimination while increasing pipeline complexity.Source: §2 · p. 2
  • Many deep-learning pCR models use images acquired during or after NACT, while multimodal inputs can hinder generalization when images or clinical variables are not standardized.The section also notes that some approaches combine multiple sequences with clinical data.Source: §2 · p. 2
  • ADC changes during NACT have been associated with pCR, but isolated baseline ADC may miss vascular or perfusion information available in DCE-MRI.The section proposes a middle ground of baseline ADC and DCE-MRI with limited clinical data.Source: §2 · p. 2

3. Materials and methods

The methods use baseline ACRIN 6698/I-SPY2 data, tumor-centered image patches, selected clinical variables, late feature fusion, and stratified five-fold cross-validation. The design limits inputs to pretreatment information and uses fold-specific procedures to reduce leakage and overfitting risk.

3.1. Dataset

  • The study used T0 data from public ACRIN 6698/I-SPY2 and retained 136 patients after modality-availability and consistency criteria.The cohort included 90 non-pCR and 46 pCR patients; DCE-MRI used a prespecified third post-contrast acquisition, whose phase choice the authors identify as a potential limitation. DCE tumor masks were reconstructed from existing annotations rather than newly drawn segmentations.Source: §3.1 · p. 2

3.2. Image preprocessing

  • For each modality, the pipeline extracted a tumor-centered 2.5D patch from three consecutive slices, stacked as 3 × 50 × 50 input channels.Non-tumor pixels were zeroed except for a fixed 5-pixel margin, balancing peritumoral information against healthy background.Source: §3.2 · pp. 2–3

3.3. Clinical variables

  • Among age, race, histological type, HR/HER2 subtype, SBR grade, and maximum diameter, only HR/HER2 subtype consistently improved AUC in preliminary testing.Numerical variables were standardized within each training partition, and categorical variables used one-hot encoding with fixed global categories.Source: §3.3 · p. 3

3.4. Multimodal architecture

  • The model uses independent EfficientNet-B0 image encoders and late fusion, combining their features with a separately processed clinical vector to predict pCR/non-pCR probabilities.Image encoders were initialized with ImageNet-pretrained weights; the clinical path and final classifier use linear layers, ReLU, and dropout.Source: §3.4 · p. 3

3.5. Training and metrics

  • Evaluation used stratified five-fold cross-validation, with preprocessing and augmentation fitted or applied within training folds to limit leakage.The epoch with the best validation AUC was selected, and inter-fold average ROC-AUC, accuracy, and F1-score were reported; external validation was absent.Source: §3.5 · pp. 3–4

4. Results

ADC alone outperformed DCE-MRI, while adding HR/HER2 subtype improved both and the ADC+DCE-MRI+HR/HER2 configuration achieved the best metrics. The interpretation compares this pretreatment result with studies using later treatment time points and notes the limited cohort and need for external validation.

4. Results

  • ADC+DCE-MRI+HR/HER2 subtype offered the best metrics, while standalone ADC achieved AUC 0.7936 versus 0.7428 for DCE-MRI.The authors describe DCE-MRI as providing complementary utility in the final multimodal configuration.Source: §4 · p. 4
  • Adding HR/HER2 subtype increased the AUC of both standalone modalities, whereas adding age, race, histological type, SBR grade, and maximum diameter did not produce average improvement.The combined ADC+DCE-MRI+HR/HER2 configuration achieved the highest AUC among the evaluated configurations.Source: §4 · p. 4

4.1. Interpretation

  • With 136 patients, small differences between cross-validation partitions may affect AUC, reinforcing the need for external validation.The authors connect this variability to the balance between multimodal complexity and overfitting risk.Source: §4.1 · p. 4
  • The model's AUC of 0.86 uses only T0 data, unlike longitudinal models that benefit from observing post-treatment changes.The authors describe ACRIN 6698 mid-treatment and post-treatment ADC AUCs of 0.60 and 0.61, respectively, and report AUCs of 0.803–0.840 for BMMR2 submissions.Source: §4.1 · p. 4
  • Among pretreatment studies, the model's AUC of 0.86 was comparable to a radiomics model using multiparametric MRI and clinical variables.The authors contrast its compact input configuration with studies using larger single-center cohorts or additional sequences and describe their own results as preliminary and internally validated.Source: §4.1 · p. 4

5. Discussion

The discussion presents pretreatment multimodal prediction as feasible, while emphasizing constraints from limited data, segmentation requirements, and the need for external validation. It also notes that HR/HER2 subtype was the only tested clinical variable to improve AUC on average.

5. Discussion

  • The ADC+DCE-MRI combination uses diffusion and enhancement/perfusion information, while HR/HER2 subtype adds biological context related to NACT response.The model does not require waiting for treatment-induced changes during NACT.Source: §5 · p. 4
  • Only HR/HER2 subtype improved AUC on average among the tested clinical variables, and selecting fewer variables may facilitate implementation.The other tested variables could introduce noise through missing values, heterogeneous categories, or inconsistent encodings.Source: §5 · p. 5
  • With 46 pCR cases in a limited cohort, each validation fold’s AUC is sensitive to individual classifications, so reported AUCs represent internal validation performance.The authors state that regularization and other training strategies do not replace external validation.Source: §5 · p. 5
  • Clinical implementation would require robust automatic or semi-automatic segmentation and evaluation of how segmentation variation affects prediction.The DCE-MRI tumor masks were reconstructed from existing dataset annotations through a radiologist-supervised process.Source: §5 · p. 5
  • External cohorts are still needed, and the added value of DCE-MRI must be weighed against gadolinium-related cost, acquisition time, and use restrictions.For contrast-free settings, ADC+HR/HER2 subtype is described as a simple and stable alternative.Source: §5 · p. 5

6. Conclusions

The conclusion reports the best cross-validation AUC for the combined ADC, DCE-MRI, and HR/HER2 subtype configuration. It recommends further external validation, interpretability work, and reliable semantic segmentation to support real-world implementation.

6. Conclusions

  • The combined ADC+DCE-MRI+HR/HER2 subtype configuration achieved the best cross-validation AUC, reaching 0.86.ADC performed better than DCE-MRI alone and when each was combined with HR/HER2 subtype.Source: §6 · p. 5
  • Future work should focus on external validation and interpretability, and test preprocessing choices to identify relevant image regions and modalities.Source: §6 · p. 5
  • Reliable automatic semantic segmentation could facilitate real-world implementation of the proposed technique as a clinical decision-support tool.Source: §6 · p. 6

Ethics statement

The study used public anonymized data from The Cancer Imaging Archive and reports that no identifiable patient information was used.

Ethics statement

Loading 2609.26105v1…