Automatic quantification of breast cancer biomarkers from multiple 18F-FDG PET image segmentation

TL;DR

This work presents a nnUNet-based pipeline to automatically segment breast tumors on 18F-FDG PET and to compute key biomarkers ($SUV_{max}$, $MTV$, $TLG$) for monitoring response after the first course of neoadjuvant chemotherapy. Baseline models trained on 243 PET_Bl scans achieve high segmentation accuracy ($DSC$ ≈ 0.89) and robust boundary delineation, while fine-tuning on 15 PET_Fu scans with active learning enables reasonable adaptation ($DSC$ ≈ 0.78) for follow-up images. The method demonstrates very strong correlation between biomarkers derived from automatic vs manual masks, with significant reductions in SUV max, MTV, and TLG after NAC (e.g., ΔSUVmax ≈ −5.22, ΔMTV ≈ −11.79 cm^3, ΔTLG ≈ −19.23 cm^3; all p<0.01). Quality-control steps and inter-observer assessments position the approach as a practical tool for automatic, objective tracking of treatment response, with potential to inform personalized treatment planning.

Abstract

Neoadjuvant chemotherapy (NAC) has become a standard clinical practice for tumor downsizing in breast cancer with 18F-FDG Positron Emission Tomography (PET). Our work aims to leverage PET imaging for the segmentation of breast lesions. The focus is on developing an automated system that accurately segments primary tumor regions and extracts key biomarkers from these areas to provide insights into the evolution of breast cancer following the first course of NAC. 243 baseline 18F-FDG PET scans (PET_Bl) and 180 follow-up 18F-FDG PET scans (PET_Fu) were acquired before and after the first course of NAC, respectively. Firstly, a deep learning-based breast tumor segmentation method was developed. The optimal baseline model (model trained on baseline exams) was fine-tuned on 15 follow-up exams and adapted using active learning to segment tumor areas in PET_Fu. The pipeline computes biomarkers such as maximum standardized uptake value (SUVmax), metabolic tumor volume (MTV), and total lesion glycolysis (TLG) to evaluate tumor evolution between PET_Fu and PET_Bl. Quality control measures were employed to exclude aberrant outliers. The nnUNet deep learning model outperformed in tumor segmentation on PET_Bl, achieved a Dice similarity coefficient (DSC) of 0.89 and a Hausdorff distance (HD) of 3.52 mm. After fine-tuning, the model demonstrated a DSC of 0.78 and a HD of 4.95 mm on PET_Fu exams. Biomarkers analysis revealed very strong correlations whatever the biomarker between manually segmented and automatically predicted regions. The significant average decrease of SUVmax, MTV and TLG were 5.22, 11.79 cm3 and 19.23 cm3, respectively. The presented approach demonstrates an automated system for breast tumor segmentation from 18F-FDG PET. Thanks to the extracted biomarkers, our method enables the automatic assessment of cancer progression.

Figures (8)

• Figure 1: Diagram of our protocol. The data from the PET$_{Bl}$ scan performed before the start of treatment are used to quantify the initial metabolism of the tumor. The data from the PET$_{Fu}$ scan performed after the first course of chemotherapy are used to assess residual metabolism of the tumor and the metabolic response compared to the results of PET$_{Bl}$. Magnetic Resonance Imaging (MRI) exam, which is an addition to the main protocol, was performed at baseline step.
• Figure 2: Axial PET images of breast tumor a before and b after first cycle of NAC for the same patient. The red arrow indicates the tumor lesion.
• Figure 3: Pipeline of the segmentation of the tumor on PET$_{Bl}$ scans. The pre-processing part explains how the input data are organized and cleaned before being fed into the network for the segmentation of PET$_{Bl}$ scans alongside the ground truth.
• Figure 4: Pipeline for the management of the PET$_{Fu}$ scans and the biomarker extraction: a Extraction of biomarkers using the segmented mask at the baseline level. b Segmentation of the PET$_{Fu}$ scans using the fine-tuned baseline model (model$_{BL}$). c Calculation of changes in the biomarkers between PET$_{Fu}$ and PET$_{Bl}$, such as SUV$_{max}$, to observe the impact of NAC. Active learning process is a process in which outliers identified in the PET$_{Fu}$ segmentation by the quality control system are then manually labeled to further refine the model. The terme "Mapping" corresponds to the extraction of the biomarkers from the region of interest associated to a segmented mask.
• Figure 5: The distribution of MTV in PET$_{Bl}$ scans, calculated from labels, is compared to the MTV ratio between PET$_{Fu}$ and PET$_{Bl}$ cases a before and b after fine-tuning. The data points encircled by a red line are identified as outliers, while those within a rectangular green boundary are considered as extreme outliers, being the farthest from the threshold. The blue perpendicular dotted line to the x-axis represents the threshold. The regression line in a and b represents the relationship between the MTV of labeled scans and the MTV ratio between PET$_{Bl}$ and PET$_{Fu}$ scans.