ConvNeXt-MIL-XGBoost: Cancer Risk Prediction Pipeline

• The paper introduces a modular pipeline that integrates a frozen ConvNeXt backbone, gated attention MIL, and XGBoost for breast cancer recurrence risk stratification from H&E WSIs.
• It employs a Transformer-inspired patch extraction and attention pooling mechanism to aggregate discriminative features across image patches.
• The study demonstrates clinical relevance by achieving 73.5% accuracy and producing interpretable attention maps to support diagnostic decisions.

ConvNeXt-MIL-XGBoost is a modular weakly-supervised deep learning pipeline designed for automated stratification of breast cancer recurrence risk from Hematoxylin and Eosin (H&E) stained whole-slide images (WSIs). Developed and benchmarked alongside CLAM-SB and ABMIL models for predicting 5-year recurrence risk tiers, the method integrates a frozen ConvNeXt-Base neural feature extractor, a gated-attention Multiple Instance Learning (MIL) aggregator, and an XGBoost gradient-boosted decision tree classifier. The pipeline leverages robust natural-image priors and interpretable feature aggregation for genomics-correlated risk prediction, with demonstrated efficacy on a dataset of 210 patient cases (Chen et al., 21 Dec 2025).

1. ConvNeXt Backbone for Patch-Level Feature Extraction

The first stage of ConvNeXt-MIL-XGBoost entails decomposing each WSI into 256×256 px tissue-containing patches, which are independently encoded using the ConvNeXt-Base convolutional neural network. The ConvNeXt-Base architecture employs a “Transformer-inspired” design characterized by large depth-wise kernels, inverted bottlenecks, and LayerNorm, facilitating multi-scale texture and context capture. Core configuration parameters include depths per stage of [3, 3, 27, 3] layers; channel dimensions of [96, 192, 384, 768]; 4×4 patch-embedding convolutions at input, and subsequent 2×2 strided convolutions between stages. Patch representations are finalized by an additional fully-connected projection to a 1024-dimensional vector, yielding approximately 90 million parameters.

ImageNet-1k pre-trained weights initialize ConvNeXt, accelerating convergence and imparting robust feature representations given the small sample size (210 WSIs). No fine-tuning is performed; the backbone serves as a frozen encoder for subsequent stages, ensuring computational efficiency and reproducibility.

2. Attention-Based Multiple Instance Learning Aggregation

Following patch encoding, the MIL framework aggregates instance-level features $x_{i,j} \in \mathbb{R}^{1024}$ into a slide-level embedding $s_i \in \mathbb{R}^{1024}$ for each slide $i$ with $N_i$ patches. The aggregation employs a gated attention pooling mechanism:

$$a_{i,j} = \mathrm{softmax}_j \left( w^T \tanh(V x_{i,j}) \odot \sigma(U x_{i,j}) \right)$$

$s_i = \sum_{j=1}^{N_i} a_{i,j} x_{i,j}$

where $V, U \in \mathbb{R}^{384 \times 1024}$ are learnable weights, $\sigma$ denotes the sigmoid function, $w \in \mathbb{R}^{384}$ is a projection vector, $\odot$ indicates element-wise multiplication, and the softmax ensures $\sum_j a_{i,j}=1$.

To address the pronounced class imbalance—evident from only 21 medium-risk slides—the attention module is trained using Focal Loss with $\gamma=2$ and disproportionately up-weighted medium-risk class ($\alpha_{medium}=3.0$, $\alpha_{low} = \alpha_{high} = 1.0$). The optimizer is Adam (β₁=0.9, β₂=0.999) with an initial learning rate of $1 \times 10^{-4}$ and batch size of 8 slides. This attention model aggregates diverse instances while focusing on discriminative tissue regions.

3. XGBoost Classification on Structured Slide Embeddings

The trained attention pooling module outputs for each WSI a 1024-dimensional embedding $s_i$, supplemented by auxiliary features including MIL logits, softmax probabilities, attention distribution summary statistics (mean, median, quartiles, skewness), and patch count $N_i$—resulting in a concatenated feature vector $f_i \in \mathbb{R}^{1047}$.

Classification utilizes XGBoost, a gradient-boosted decision tree algorithm, targeting slide-level three-tier risk prediction ($y_i\in\{0,1,2\}$). The multi-class softmax objective combines cross-entropy loss with tree complexity penalties:

$$L(\Theta) = \sum_{i=1}^{M} \ell(y_i, \hat{y}_i^{(t)}) + \sum_{k=1}^{T} \Omega(f_k)$$

$$\Omega(f_k) = \gamma T_k + \frac{1}{2}\lambda\sum_j w_{k,j}^2$$

with $T=200$ trees, learning rate $\eta=0.10$, maximum tree depth 6, $\gamma=1.0$, and L₂ regularization $\lambda=1.0$. Hyperparameters are tuned on validation splits, and the XGBoost optimizer applies second-order gradient methods.

4. Training Protocol and Data Management

ConvNeXt-MIL-XGBoost training follows a robust 5-fold cross-validation procedure, with data splits stratified by risk class to preserve distributional balance. For each fold, 80% of slides are allocated for backbone and attention training, 10% for MIL validation and early stopping, and 10% as an unseen test set.

Risk labels are based on consensus between the 21-gene Recurrence Score and clinicopathological adjudication. MIL model training employs Focal Loss, while XGBoost operates on multi-class logistic loss. Adam optimizer (attention module) and XGBoost’s native optimizer ensure efficient convergence. No data augmentation is utilized beyond randomized patch ordering; patch segmentation and extraction are deterministic.

5. Comparative Performance and Ablation Insights

ConvNeXt-MIL-XGBoost exhibits mean classification accuracy of $73.5\% \pm 3.8\%$ across five cross-validation folds. Comparative results are summarized:

Model	Mean AUC	Mean Accuracy
CLAM-SB	$0.836 \pm 0.062$	$76.2\% \pm 4.5\%$
ABMIL	$0.767 \pm 0.046$	$70.9\% \pm 5.1\%$
ConvNeXt-MIL-XGBoost	–	$73.5\% \pm 3.8\%$

No AUC was computed for the XGBoost pipeline, and formal significance testing was omitted. Notably, ConvNeXt-Base provides superior patch embeddings over ResNet-18, raising classification accuracy by ≈5%. The gated attention (Sigmoid×tanh) improves aggregation over mean pooling (+4% accuracy). XGBoost outperforms equivalent multi-layer perceptron classifiers (same $f_i$ input), with MLP achieving only 70.0% accuracy and exhibiting greater hyperparameter sensitivity.

6. Clinical Deployment Considerations

The ConvNeXt-MIL-XGBoost framework supports integration as a decision-support module in digital pathology environments. Automated processing of digitized slides can provide rapid three-tier recurrence risk scores and deploy attention heatmap overlays to highlight diagnostically relevant tissue regions. This workflow can prioritize high-risk cases for review and assist in adjudicating borderline cases.

Recommended deployment steps include multi-center prospective validation, regulatory approval, and integration with laboratory information systems. Ongoing monitoring of model drift due to factors such as staining or scanner variations and re-calibration with local cohorts are advised to sustain clinical efficacy.

7. Modular Architecture and Interpretability

ConvNeXt-MIL-XGBoost is structured to decouple representation learning (ConvNeXt backbone), instance aggregation (attention MIL), and structured classification (XGBoost). This modularity yields an interpretable pipeline for genomics-correlated risk prediction, supporting transparent diagnostic audits and facilitating adaptation to evolving clinical requirements. The model’s attention maps and multi-tier output structure bolster its utility within computational pathology workflows, promoting rapid, cost-effective clinical decision support while maintaining methodological robustness (Chen et al., 21 Dec 2025).