Viable tumor cell density after neoadjuvant chemotherapy assessed using deep learning model reflects the prognosis of osteosarcoma

Abstract

Prognosis after neoadjuvant chemotherapy (NAC) for osteosarcoma is generally predicted using manual necrosis-rate assessments; however, necrosis rates obtained in these assessments are not reproducible and do not adequately reflect individual cell responses. We aimed to investigate whether viable tumor cell density assessed using a deep-learning model (DLM) reflects the prognosis of osteosarcoma. Seventy-one patients were included in this study. Initially, the DLM was trained to detect viable tumor cells, following which it calculated their density. Patients were stratified into high and low-viable tumor cell density groups based on DLM measurements, and survival analysis was performed to evaluate disease-specific survival and metastasis-free survival (DSS and MFS). The high viable tumor cell density group exhibited worse DSS (p = 0.023) and MFS (p = 0.033). DLM-evaluated viable density showed correct stratification of prognosis groups. Therefore, this evaluation method may enable precise stratification of the prognosis in osteosarcoma patients treated with NAC.

Fig. 1. Overview of the study and patient selection.

This study comprises two phases. Phase 1 is intended to develop and evaluate the DLM, which detects viable tumor cells in pathological images and the validation for the established DLM using the specimen from the other hospital. Phase 2 aims to perform survival analysis based on viable tumor cell density. WSI whole-slide images, DLM deep-learning model.

Fig. 2. Schematic overview of Phase 1.

a The 15 patients were divided into five subsets of 3 each. Each subset contained one patient with only biopsy specimens and two with biopsy and resection specimens. Three pathologists selected one patch-extracted image (1024 × 1024 pixel) at 400× magnification per WSI and annotated the nuclei of viable tumor cells by consensus. Thus, each subset had five patch-extracted images, for a total of 25 images. b Five-fold cross-validation was performed, with three subsets used for training, one subset for validation, and one subset for testing. DLM deep-learning model, WSI whole-slide images.

Fig. 3. Calculation workflow of the viable tumor cell density determination in Phase 2.

The resected tumor was sliced (Step 1), and the sliced specimens were scanned as WSIs. The pathologist annotated the resected tumor’s margins in all WSIs (Step 2). Next, patch-extracted images were generated from inside the annotated tumor area of the WSI (Step 3). The trained DLM detected viable tumor cells inside the tumor in patch-extracted images (Step 4), and counted the viable tumor cells (Step 5). The area inside the annotated tumor border in the WSI was calculated (Step 6), and the area of the tumor (mm²) (“Area”) and the total number of detected viable tumor cells (“Sum”) were evaluated (Step 7). Finally, the viable cell density defined as “Sum∕Arⅇa” was calculated. DLM deep-learning model.

Fig. 4. Survival analysis based on viable tumor cell density and necrosis rate.

a, b Kaplan–Meier plots and statistical test results for DSS and MFS on viable tumor cell density (cut-off value 400/mm²). c, d: The same analyses on manually evaluated necrosis rate. DLM deep-learning model, DSS disease-specific survival, MFS metastasis-free survival.

Fig. 5. Comparison between viable cell density by DLM and manually determined necrosis rate.

a Case distribution by DLM and pathologist evaluation group. Red background squares indicate the number of cases in which both the DLM and pathologists judged the prognosis as poor, and blue indicates cases in which both judged the prognosis as good. The yellow represents the number of cases in which the DLM and the pathologists were divided in their decisions. b–d Example of the detection result at Phase 2. Each red circle denotes a detected viable tumor cell by DLM. DLM deep-learning model.