Vision Transformers for Preoperative CT-Based Prediction of Histopathologic Chemotherapy Response Score in High-Grade Serous Ovarian Carcinoma

Francesca Fati; Felipe Coutinho; Marika Reinius; Marina Rosanu; Gabriel Funingana; Luigi De Vitis; Gabriella Schivardi; Hannah Clayton

cs.CVApr 10, 2026

Vision Transformers for Preoperative CT-Based Prediction of Histopathologic Chemotherapy Response Score in High-Grade Serous Ovarian Carcinoma

Authors: Francesca Fati, Felipe Coutinho, Marika Reinius, Marina Rosanu, Gabriel Funingana, Luigi De Vitis, Gabriella Schivardi, Hannah Clayton,

Organizations: Department of Gynecologic Oncology, European Institute of Oncology, IEO, IRCCS, Italy. · Department of Electronics, Information and Bioengineering, Politecnico di Milano, Italy. · INESC TEC, Faculty of Engineering, University of Porto, Portugal. · Department of Oncology, University of Cambridge, United Kingdom. · Cancer Research UK Cambridge Centre, United Kingdom. · Cancer Research UK Cambridge Institute, United Kingdom. · Cambridge University Hospitals NHS Foundation Trust, United Kingdom. · Department of Obstetrics and Gynecology, Mayo Clinic, USA. · Early Cancer Institute, University of Cambridge, United Kingdom. · Research Center for Medical Image Analysis and AI, Danube Private University, Austria. · Department of Pathology, European Institute of Oncology, IEO, IRCCS, Italy. · Department of Oncology and Hemato-Oncology, University of Milan, Italy. · Department of Radiologic Sciences, Universit`a Cattolica del Sacro Cuore, Italy. · Institute of Biomedical Engineering, Karlsruhe Institute of Technology, Germany. · Department of Radiology, Mayo Clinic, USA.

Abstract

Purpose. High-grade serous ovarian carcinoma (HGSOC) is characterized by pronounced biological and spatial heterogeneity and is frequently diagnosed at an advanced stage. Neoadjuvant chemotherapy (NACT) followed by delayed primary surgery is commonly employed in patients unsuitable for primary cytoreduction. The Chemotherapy Response Score (CRS) is a validated histopathological biomarker of response to NACT, but it is only available postoperatively. In this study, we investigate whether pre-treatment computed tomography (CT) imaging and clinical data can be used to predict CRS as an investigational decision-support adjunct to inform multidisciplinary team (MDT) discussions regarding expected treatment response. Methods. We proposed a 2.5D multimodal deep learning framework that processes lesion-dense omental slices using a pre-trained Vision Transformer encoder and integrates the resulting visual representations with clinical variables through an intermediate fusion module to predict CRS. Results. Our multimodal model, integrating imaging and clinical data, achieved a ROC-AUC of 0.95 alongside 95% accuracy and 80% precision on the internal test cohort (IEO, n=41 patients). On the external test set (OV04, n=70 patients), it achieved a ROC-AUC of 0.68, alongside 67% accuracy and 75% precision. Conclusion. These preliminary results demonstrate the feasibility of transformer-based deep learning for preoperative prediction of CRS in HGSOC using routine clinical data and CT imaging. As an investigational, pre-treatment decision-support tool, this approach may assist MDT discussions by providing early, non-invasive estimates of treatment response.

Figures & tables

Figure 1: Schematic of the proposed 2.5D multimodal classification pipeline. The three axial slices with the highest lesion density are extracted from the segmented omentum and stacked to form a multi-channel input for a DINOv3-small encoder. The resulting visual representation is combined with encoded clinical variables (age, CA125) via an intermediate fusion layer. The final MLP head outputs the probability of a non- or partial pathological chemotherapy response (CRS 1–2) versus a complete response (CRS 3).

Figure 2: Representative pre- and post-treatment CT scans of HGSOC patients from the IEO and OV04 cohorts, stratified by CRS grade: CRS 1 (minimal or no response), CRS 2 (partial response), and CRS 3 (complete or near-complete response). Axial, sagittal, and coronal views are shown for each case, with omental tumor regions used for CRS assessment highlighted in blue on pre-treatment scans and in violet on post-treatment scans.

Characteristic	IEO (Internal)			OV04 (External)
	Train (n=191)	Validation (n=39)	Test (n=41)	Test (n=70)
Age (years)	64 [54–70]	62 [55–66]	65 [55–74]	63 [54–70]
CA-125 (U/mL)	1192 [404–2896]	1006 [345–1722]	914 [472–2875]	898 [286–2070]
CRS, n (%)
CRS 1 (no/minimal)	80 (41.7%)	17 (43.6%)	17 (41.5%)	19 (27.1%)
CRS 2 (partial)	86 (44.8%)	18 (46.1%)	18 (43.9%)	21 (30.0%)

Table 1: Baseline clinical characteristics and chemotherapy response score (CRS) distribution of the IEO and OV04 cohorts. The table reports patient age and serum cancer antigen 125 (CA-125) levels at diagnosis, together with CRS labels, for the IEO cohort (subdivided into training, validation, and internal test sets) and the independent external OV04 cohort. Continuous variables are presented as median [IQR].

	IEO (Internal)			OV04 (External)
	CRS1 (n=114)	CRS2 (n=122)	CRS3 (n=38)	CRS1 (n=19)	CRS2 (n=21)	CRS3 (n=30)
Age (years)	65.5	63.5	60.5	64.2	65.5	61.0
CA-125 (U/mL)	655.9	946.7	426.5	867.0	584.0	1423.5
Tumor Morphology
Volume [ $cm^{3}$ ]	102.4	80.98	15.62	119.2	61.04	36.6
Surface Area [ $cm^{2}$ ]	468.6	437.8	167.4	1276	781	457.9

Table 2: Comparison of patient demographics and omental tumor morphology across CRS grades in the IEO internal validation set and the OV04 external validation set. Data are presented as median values for age, pre-treatment CA-125 levels, and quantitative tumor morphology metrics, including tumor volume, surface area, and the largest connected component (CC) fraction. Patients are stratified by CRS classification: CRS 1 (minimal or no response), CRS 2 (partial response), and CRS 3 (complete or near-complete response).

Panel of Table 3 — Table 3: (a) Discriminative performance across external cohorts (ROC-AUC, F1-score, precision, recall, accuracy). (b) Corresponding confusion matrix components (percentages relative to positives/negatives: TP, TN, FP, FN). All values are medians with 95% confidence intervals, evaluated at ( $\tau_{0.5}$ and $\tau_{opt}=0.69$ ).

				IEO Test	OV04
CT	Volume	Age	CA125	AUC	AUC
✗	✓	✓	✓	0.644 [0.459–0.792]	0.537 [0.407–0.674]
✓	✗	✗	✗	0.894 [0.727–1.000]	0.469 [0.312–0.632]
✓	✗	✓	✗	0.811 [0.333–1.000]	0.572 [0.414–0.727]
✓	✗	✗	✓	0.933 [0.744–1.000]	0.521 [0.364–0.677]
✓	✗	✓	✓	0.952 [0.862–1.000]	0.677 [0.527–0.806]

Table 4: Median AUC (95% CI) for each feature configuration in the ablation study, evaluated on the IEO test split and the external OV04 cohort. Each row indicates which features are included: CT imaging, lesion volume, age, and CA-125 levels (✓ = included, ✗ = excluded). AUC values are reported at the threshold maximizing the F1 score for each cohort.

Panel of Figure 4 — Figure 4: ROC curves for the ablation study on (a) IEO and (b) OV04 cohorts. The full multimodal model (CT + Age + CA-125) achieves the highest AUC, with the largest gains on OV04 observed when clinical features are added.

Clinical Outcome Prediction

Clinical outcome prediction uses patient records, imaging, signals, and biomarkers to estimate risks, disease trajectories, treatment response, and survival.

See all 265 papers

Predicting Response to Neoadjuvant Chemotherapy in Ovarian Cancer from CT Baseline Using Multi-Loss Deep Learning
May 14, 2026Francesco Pastori, Francesca Fati, Marina Rosanu +8Clinical Outcome Prediction Oncology
Department of Gynecologic Oncology, European Institute of Oncology, IEO, IRCCS, Milan, Italy · Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milan, Italy · Department of Obstetrics and Gynecology, Mayo Clinic, Rochester, USA +2
Neoadjuvant chemotherapy response prediction using pretreatment diffusion and contrast-enhanced magnetic resonance imaging with clinical variables
Aug 18, 2026Pablo García Marcos, Paula Puerta González, Guillermo Lorenzo +6Cancer Drug Response Prediction Clinical Outcome Prediction
Computer Science Department, University of Oviedo, Asturias, Spain · Biomedical Engineering Center (BME), University of Oviedo, Asturias, Spain · Electrical Engineering Department, University of Oviedo, Asturias, Spain +7
Transcriptome-informed multi-modal AI for predicting neoadjuvant therapy response from breast cancer biopsies
Oct 2, 2026Jungkyu Park, Dhruva Biswas, Joseph Cappadona +22Clinical Outcome Prediction Computational Pathology
Ataraxis AI, New York, NY, USA. · Breast Oncology Program, Dana-Farber Cancer Institute, Boston, MA, USA. · Harvard Medical School, Boston, MA, USA. +13

May 14, 2026J/K move · Enter open · S save

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

Francesco Pastori, Francesca Fati, Marina Rosanu, Luigi De Vitis+7

Department of Gynecologic Oncology, European Institute of Oncology, IEO, IRCCS, Milan, Italy · Department of Electronics, Information and Bioengineering, Politecnico di Milano, Milan, Italy · Department of Obstetrics and Gynecology, Mayo Clinic, Rochester, USA+2

Clinical Outcome Prediction Oncology

Open paper AskPDF arXiv

Ovarian cancer is the most lethal gynecologic malignancy: around 60% of patients are diagnosed at an advanced stage, with an associated 5-year survival rate of about 30%. Early identification of non-responders to neoadjuvant chemotherapy remains a key unmet need, as it could prevent ineffective therapy and avoid delays in optimal surgical management. This work proposes a non-invasive deep learning framework to predict neoadjuvant chemotherapy response from pre-treatment contrast-enhanced CT by leveraging automatically derived 3D lesion masks. The approach encodes axial slices with a partially fine-tuned pretrained image encoder and aggregates slice-level representations into a volumetric embedding through an attention-based module. Training combines classification loss with supervised contrastive regularization and hard-negative mining to improve separation between ambiguous responders and non-responders. The method was developed on a retrospective single-center cohort from the European Institute of Oncology (Milan, IT), including 280 eligible patients (147 responder, 133 non-responder). On the test cohort, the 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). Overall, these results suggest that the proposed architecture learns clinically relevant predictive patterns and provides a robust foundation for an imaging-based stratification tool.

Abstract

Clinical Outcome Prediction

Explore similar work

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

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

Transcriptome-informed multi-modal AI for predicting neoadjuvant therapy response from breast cancer biopsies