💻 computer science

Predicting Response to Neoadjuvant Chemotherapy in Ovarian Cancer from CT Baseline Using Multi-Loss Deep Learning

This paper presents a multi-loss deep learning framework that utilizes pre-treatment contrast-enhanced CT scans and automatically derived 3D lesion masks to predict neoadjuvant chemotherapy response in ovarian cancer patients, achieving a ROC-AUC of 0.73 on a retrospective cohort of 280 patients.

Original authors: Francesco Pastori, Francesca Fati, Marina Rosanu, Luigi De Vitis, Lucia Ribero, Gabriella Schivardi, Giovanni Damiano Aletti, Nicoletta Colombo, Jvan Casarin, Francesco Multinu, Elena De Momi

Published 2026-05-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Francesco Pastori, Francesca Fati, Marina Rosanu, Luigi De Vitis, Lucia Ribero, Gabriella Schivardi, Giovanni Damiano Aletti, Nicoletta Colombo, Jvan Casarin, Francesco Multinu, Elena De Momi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a patient is diagnosed with ovarian cancer, a serious condition where time is often the most critical factor. Doctors have two main paths: operate immediately or give the patient chemotherapy first (neoadjuvant chemotherapy) to shrink the tumor before surgery. The problem is that about 40% of patients don't actually respond to this chemotherapy. If they don't respond, they are wasting precious time on a treatment that isn't working, delaying the surgery they actually need.

This paper presents a new "digital detective" designed to solve this problem before treatment even begins.

The Detective's Toolkit: The CT Scan

Instead of asking a doctor to look at a 3D CT scan (a detailed 3D X-ray) with their eyes, the researchers built a computer program. Think of the CT scan as a loaf of bread. The computer doesn't just look at the whole loaf; it slices it up into individual pieces (axial slices) to examine the texture of the "dough" (the tumor) in detail.

How the "Digital Detective" Learns

The system uses a special type of artificial intelligence called a Vision Transformer. You can think of this as a very smart student who has already read thousands of books about general shapes and patterns (pre-trained on other data). Now, this student is being taught specifically about ovarian tumors.

Here is how the training works, using a creative analogy:

The Mask (The Outline): First, the computer automatically draws a tight outline around the tumor in every slice of the CT scan, ignoring the healthy tissue. It's like tracing the tumor with a highlighter.
The Two-Headed Teacher: The computer learns using a special "multi-loss" strategy, which is like a teacher using two different methods to ensure the student really understands:
- Method A (The Quiz): The teacher asks, "Is this patient a responder or a non-responder?" and checks the answer. This is the standard way to learn.
- Method B (The Sorting Game): This is the clever part. The teacher takes two patients who both responded to treatment and says, "These two are similar; keep them close together in your mind." Then, the teacher takes a patient who responded and one who didn't, and says, "These two are different; push them far apart."
- The "Hard" Lesson: Sometimes, the teacher picks two patients who look almost identical on the scan but had different outcomes. The computer is forced to work extra hard to find the tiny, invisible difference between them. This is called "hard-negative mining." It's like training a wine taster to distinguish between two bottles of wine that look exactly the same but taste slightly different.

The Results: Did the Detective Get It Right?

The researchers tested this system on 280 patients from a hospital in Milan. They split the group: they taught the system on 226 patients and then tested it on 54 new patients it had never seen before.

The Score: The system achieved a score (called ROC-AUC) of 0.73. In the world of medical prediction, this is a solid result, meaning the computer is significantly better than random guessing and better than some older methods they compared it against.
The F1-Score: It got a score of 0.70, which balances how often it was right with how often it caught the cases it needed to catch.

What the Computer "Saw"

The researchers didn't just trust the number; they peeked under the hood to see what the computer was looking at.

Focus on the Core: When the computer made a correct prediction, it paid the most attention to the slices in the middle of the tumor (the "core"), where the tumor is thickest. It ignored the edges, just like a human expert would.
Confusion: When the computer made a mistake, its attention was scattered and confused, looking at the edges as much as the center.
The "Borderline" Problem: The researchers found that when the computer was unsure, it was often because the patient was right on the edge of the definition of "responder" vs. "non-responder." It wasn't that the computer was broken; it was that the medical definition itself was blurry for those specific cases.

The Bottom Line

This paper claims that by using a smart AI that looks at the 3D shape of the tumor and learns through a special "sorting game," we can predict with reasonable accuracy whether a patient will respond to chemotherapy before they even start.

The authors state that this provides a "robust foundation" for a tool that could help doctors decide earlier if a patient needs a different strategy, potentially saving time and avoiding ineffective treatments. They emphasize that the system learns patterns that are "clinically relevant," meaning it's not just finding random noise, but actual signs that matter for patient care.

Technical Summary: Predicting Response to Neoadjuvant Chemotherapy in Ovarian Cancer from CT Baseline Using Multi-Loss Deep Learning

Problem Statement
Ovarian cancer remains the most lethal gynecologic malignancy, with approximately 60% of patients diagnosed at an advanced stage. Standard management often involves neoadjuvant chemotherapy (NACT) followed by interval debulking surgery. However, a critical unmet clinical need exists: roughly 40% of patients assigned to NACT fail to derive objective benefit from the therapy. Identifying these non-responders prior to treatment initiation is essential to prevent ineffective therapy and avoid delays in optimal surgical management. Current predictive tools often rely on clinical baselines or radiomic features that may lack robustness or require manual segmentation, necessitating a non-invasive, automated solution capable of analyzing pre-treatment imaging to stratify patients effectively.

Methodology
The authors propose a non-invasive deep learning framework that predicts NACT response using pre-treatment contrast-enhanced 3D CT scans. The pipeline processes volumetric data ( $V_{CT}$ ) through the following stages:

Lesion Segmentation: An open-source deep learning segmentation model ( $M_S$ ) automatically generates 3D lesion masks from the raw CT volumes, isolating tumor regions.
Slice Encoding: The 3D mask volume is reshaped into a sequence of axial slices. Each slice is converted into a 3-channel representation and processed independently by a pre-trained Vision Transformer (ViT) encoder ( $E_{ViT}$ ), specifically the DINOv3 Base Patch16. The encoder extracts token embeddings, from which the classification token ( $t_{cls}$ ) serves as a global descriptor for each slice.
Volumetric Aggregation: The classification tokens from all slices are concatenated and augmented with learnable positional embeddings to preserve slice ordering. An attention-based pooling module aggregates these slice-level representations, modeling inter-slice relationships, followed by mean pooling to produce a single volumetric embedding ( $z_V$ ).
Multi-Head Prediction: The shared embedding $z_V$ $z_{V}$ branches into two heads:
- Classification Head: A Multi-Layer Perceptron (MLP) that outputs binary logits (responder vs. non-responder).
- Projection Head: An MLP that maps the embedding to a low-dimensional space ( $z_p$ ) for contrastive learning.
Training Strategy: The model employs a transfer learning approach where the ViT encoder is fine-tuned using LoRA (Low-Rank Adaptation) adapters restricted to the last Transformer block to prevent overfitting on the small dataset. The training objective combines two losses:
- Cross-Entropy Loss ( $L_{CE}$ ): Standard classification loss.
- Supervised Contrastive Loss ( $L_{SupCon}$ ): A margin-aware loss that pulls embeddings of the same class closer and pushes different classes apart by a margin $m$ .
- Hard Negative Mining: During training, negative pairs are dynamically selected based on their proximity in the embedding space to focus optimization on ambiguous cases. The weight of the contrastive loss ( $\alpha$ ) is linearly increased from 0 to 0.3 during training to stabilize initial representation learning before enforcing geometric constraints.

Key Contributions

Architecture: Integration of a pre-trained ViT with an attention-based volumetric aggregation module to process 3D lesion masks without requiring manual feature engineering.
Training Objective: A multi-loss framework combining cross-entropy with supervised contrastive learning and hard-negative mining to improve the separation of ambiguous responders and non-responders.
Efficiency: Utilization of LoRA for parameter-efficient fine-tuning, enabling the adaptation of large pre-trained models to a specific medical domain with limited data (280 patients).
Ablation Analysis: Systematic comparison demonstrating that both the contrastive loss and the fine-tuning strategy are critical for performance, outperforming models using only cross-entropy or frozen backbones.

Results
The model was evaluated on a retrospective single-center cohort from the European Institute of Oncology, comprising 280 eligible patients (147 responders, 133 non-responders), split into a development set (226) and an independent test set (54).

Performance: On the test set, the proposed Multi-Loss Fine-Tuned (ML-FT) model achieved a ROC-AUC of 0.73 (95% CI: 0.58–0.86) and an F1-score of 0.70 (95% CI: 0.56–0.82).
Comparative Analysis: The ML-FT model significantly outperformed:
- A baseline re-implementation of Yin et al. (AUC 0.47).
- A variant without contrastive loss (CE-FT, AUC 0.66).
- A variant with a frozen backbone (ML-FR, AUC 0.59).
Interpretability: Attention weight analysis revealed that the model prioritizes central slices containing the lesion core, aligning with clinical intuition that central tumor cross-sections provide the most discriminative morphological information.
Calibration: The model exhibited systematic under-confidence for the positive class, suggesting conservative probability estimates.
Error Analysis: Uncertain but correct predictions were often found near the intrinsic RECIST 1.1 cutoff, indicating that some ambiguity stems from the reference standard itself rather than model failure.

Significance and Claims
The paper claims that the proposed architecture successfully learns clinically relevant predictive patterns from pre-treatment lesion morphology alone. The results suggest that the combination of volumetric ViT encoding and supervised contrastive learning provides a robust foundation for an imaging-based stratification tool. By effectively distinguishing between responders and non-responders, the tool has the potential to support personalized treatment planning, reduce ineffective chemotherapy exposure, and facilitate timely surgical management. The authors emphasize that the model's ability to handle ambiguous cases through contrastive regularization addresses a key limitation of standard classification approaches in this domain.