Sarcomas are cancers arising from connective tissues such as muscle, bone, and fat. They are graded using the FNCLCC system (French Federation of Cancer Centers Sarcoma Group), which classifies tumor aggressiveness based on how differentiated the cells appear, the extent of tissue death within the tumor, and the mitotic rate -- how rapidly cells are dividing.
In parallel with the mitotic count, the Ki-67 proliferative index (PI) is increasingly used as a biomarker in sarcoma assessment. Ki-67 is a protein present only in actively dividing cells and is detected using immunohistochemistry (IHC) -- a staining technique that marks Ki-67-positive nuclei with a brown color while leaving non-dividing cells stained blue. The proportion of brown-stained nuclei to total nuclei within the highest-activity region of a tumor is the Ki-67 PI.
Manually calculating this ratio is tedious and subjective. Pathologists visually scan slides for hotspot regions -- areas with the densest tumor cell activity -- and estimate the percentage of stained cells. This process is slow, prone to inter-observer variability (different pathologists reaching different conclusions on the same slide), and affected by fatigue and time pressure. For a biomarker used to guide treatment decisions, such inconsistency carries real clinical risk.
The study's central goal was to develop an AI-based tool capable of consistently counting Ki-67-positive nuclei within any specified region, then test whether providing pathologists with AI output would bring their independent assessments into closer agreement.
The AI system operated on digitized whole slide images (WSIs) processed through a two-channel input: the Haematoxylin channel, which captures the blue staining of all nuclei, and the DAB (diaminobenzidine) channel, which captures the brown staining present only in Ki-67-positive nuclei. Separating these channels allows the algorithm to independently detect all nuclei and then determine which are positive.
The core of the approach was the watershed algorithm -- a classical computer vision technique that treats an image like a topographic landscape. It first applies intensity thresholding to separate foreground (nuclei) from background (empty tissue). Then a distance transformation calculates how far each pixel is from the nearest background region, creating a distance map where peaks correspond to the centers of individual nuclei. These peaks serve as seeds that anchor the watershed segmentation, allowing it to delineate each nucleus individually even when nuclei are tightly packed.
Once nucleus boundaries were established, the DAB staining intensity within each segmented nucleus was measured. Nuclei exceeding a staining intensity threshold were classified as Ki-67 positive; all others were classified as negative. The final Ki-67 PI for each region of interest was the ratio of positive to total nuclei, expressed as a percentage. This entire process ran automatically once a pathologist defined the region boundaries.
Key calibration steps were required to handle sarcoma-specific challenges. Sarcomas can contain highly pleomorphic cells -- abnormally large, irregularly shaped nuclei -- which the algorithm initially misidentified as multiple overlapping cells. This was corrected during calibration. The algorithm's parameters were fine-tuned using 15 calibration ROIs that were kept separate from the 440 ROIs used in the main validation study.
The study used 88 sequential sarcoma cases sourced from archival material at Singapore General Hospital, including both needle biopsy specimens and larger surgical resection specimens. The original diagnoses and previously recorded Ki-67 scores were not retrieved, ensuring that pathologist assessments were made independently of prior knowledge. Five random regions of interest were drawn on each case by in-house annotators, yielding a total of 440 ROIs with areas ranging from 0.02 mm2 to 3.15 mm2.
The evaluation was structured as a two-phase study. In Phase 1 (no AI), three pathologists independently scored each ROI by visual inspection, estimating Ki-67 PI at 5 to 10 percent intervals. After completing all 440 ROIs, a mandatory two-week washout period was enforced to prevent memory of previous scores from influencing the second assessment.
In Phase 2 (with AI), pathologists assessed the same 440 ROIs again, this time with the AI-generated Ki-67 PI and cell segmentation overlays visible. The pathologists could accept the AI's output or adjust it based on their clinical judgment. The degree to which each pathologist's Phase 2 score differed from their Phase 1 score was used to measure how much the AI influenced their assessments and reduced their individual variability.
Agreement between pathologists was quantified using root mean squared error (RMSE) -- a statistical measure that captures the average magnitude of scoring differences between any two pathologists. Bland-Altman plots were used to visualize agreement, plotting the difference between two methods against their average value, with 95% confidence intervals shown as dashed lines.
Before AI assistance, the three pathologists showed substantial disagreement with each other. The RMSE between Pathologist 1 and Pathologist 2 was 18.10; between Pathologist 2 and 3 it was 12.12; and between Pathologist 1 and 3 it was 12.01. The average inter-pathologist discordance without AI was 14.08 RMSE. In practical terms, this means that for any given ROI, pathologists regularly differed by 10 to 20 percentage points in their Ki-67 estimates.
After AI assistance, these differences collapsed dramatically. The RMSE between Pathologist 1 and 2 fell to 2.53; between Pathologist 2 and 3 it fell to 2.93; and between Pathologist 1 and 3 it fell to 2.10. The average inter-pathologist discordance with AI was 2.52 RMSE -- an 82.1% reduction in disagreement. The Bland-Altman plots confirmed this visually: the wide scatter between pathologist pairs in Phase 1 condensed to a narrow band centered near zero in Phase 2.
Intra-observer changes were also substantial but varied between individuals. Pathologist 2's scores shifted by within 7.5% after viewing AI outputs, suggesting their manual assessments were already close to the algorithm. Pathologist 1 shifted by up to 30% and Pathologist 3 by up to 57.5%, with the largest single adjustment being a correction from 65% to 7.5% PI on one ROI -- the pathologist had visually overestimated the proportion of positive cells by a wide margin before seeing the count.
Pathologists agreed with the AI's output on 93.5% of ROIs on average (416 out of 440 for Pathologist 1, 415 for Pathologist 2, 404 for Pathologist 3). When disagreements occurred, the primary cause was background debris in the tissue being falsely detected as positive nuclei by the AI, an error concentrated in ROIs with low Ki-67 PI values between 0 and 10 percent.
Most prior research on AI-assisted Ki-67 scoring has focused on breast cancer, where the biomarker is more established as a clinical decision point. Studies in that field showed that digital evaluation outperforms visual estimation, with one large study of 90 pathologists assessing 10 breast cancer tissue microarrays reporting an 11.9% reduction in turnaround time and 84% pathologist confidence in the AI tool. Studies in prostate and sarcoma tissue types have been far less common, making the current work one of the first to rigorously evaluate AI Ki-67 scoring in soft tissue sarcomas with their characteristic morphological complexity.
The sarcoma context posed unique challenges. Sarcomas can contain cells with highly irregular and enlarged nuclei -- so-called pleomorphic cells -- which the algorithm initially split into multiple smaller nuclei, artificially inflating the cell count. Calibration resolved this, but it illustrates why AI tools trained primarily on one tissue type cannot simply be transplanted to another without validation and recalibration.
The 6.5% of cases where pathologists overrode the AI were concentrated in ROIs with very low Ki-67 PI (0 to 10 percent). In these regions, even a small amount of background debris -- tissue fragments, staining artifacts, or inflammatory cells -- falsely flagged as positive can cause the AI to significantly overestimate the true PI. This highlights that AI-assisted scoring still requires pathologist oversight rather than passive acceptance of the algorithm's output.
A fundamental limitation that the current AI version cannot overcome is the inability to distinguish tumor cells from surrounding normal tissue or inflammation. This is why pathologists were required to manually draw ROI boundaries before the AI could run. The AI operated only within pre-specified areas; autonomous hotspot detection and tumor demarcation are capabilities the authors intend to develop in subsequent versions of the tool.
The study's core message is that AI-assisted Ki-67 scoring functions best as a concordance tool rather than a replacement for pathological judgment. By providing an objective cell count, the AI anchors pathologist assessments around a common reference point, substantially reducing the variability that arises from individual estimation habits, fatigue, and subjective threshold calibration. The combination of pathologist expertise for ROI selection and AI precision for nuclear counting produced far more consistent results than either alone.
The remaining human element -- ROI boundary delineation -- is not trivial. Different pathologists may select different hotspot areas from the same slide, which means some inter-observer variability is preserved even when AI is used for counting within agreed-upon regions. Future development should incorporate automated hotspot detection to remove this source of variability as well.
For broader clinical adoption, the study recommends validation at multiple institutions given that this work drew entirely from Singapore General Hospital. Staining protocols, scanner characteristics, and reporting conventions differ between centers, and a tool calibrated for one institution's slides may perform differently elsewhere. The authors propose testing across at least two to three institutions as the minimum standard before widespread clinical deployment.