A critical accuracy gap exists. Deep learning networks designed to classify cells in histological tissue slides have shown unsatisfyingly low accuracy, typically below 80%, making them inadequate for clinical diagnostic use.
Trained pathologists are estimated to correctly classify characteristic cells such as tumor cells with 98 to 100 percent accuracy. Existing AI tools fall far short of this benchmark, limiting their clinical adoption.
Commercial apps relying on cell classification for tasks like PD-L1 scoring or counting cell divisions have struggled to prove they are more efficient or reliable than standard human interpretation in practice.
The researchers encountered this problem firsthand, finding that even after training on more than 100,000 cells across various setups, their networks plateaued at 80 to 85 percent accuracy due to variation in tissue processing and cell biology across different slides.
The core innovation is case-by-case adaptation. Rather than training a single universal model, researchers established a baseline convolutional neural network and then fine-tuned it for each individual tumor case by adding a small number of representative cells from that specific slide.
The technique uses the HoVer-Net segmentation architecture as its foundation, a framework originally developed for simultaneous nucleus segmentation and classification. The network was pretrained on a limited set of carefully selected typical cells before being deployed on new cases.
For each new lung adenocarcinoma slide, the pathologist examines the network's predictions and adds 5, 10, or 20 cells from that specific tumor to the training data. The network then retrains and its performance is re-evaluated, mimicking the way a clinician would calibrate their judgment to a particular tissue sample.
This loop continues until the network reaches an acceptable true positive rate for classifying tumor cells, making the process interactive, transparent, and verifiable rather than a black-box prediction.
Thirty in-house lung adenocarcinoma cases were split evenly. Fifteen tumors provided training data for a baseline network, while the remaining fifteen were reserved for testing how adding 0, 5, 10, or 20 additional cells from each tested tumor affected classification accuracy.
The baseline network was trained with 300 mesenchymal cells, 300 lung adenocarcinoma tumor cells, 282 eosinophilic granulocytes, and 300 lymphocytes, covering the most common cell types found in lung tissue slides. Training ran for 200 epochs with an 80/20 validation split.
Performance was evaluated on single high-power fields initially, then extended to three high-power fields per tumor representing areas of low, medium, and high tumor cell density to assess generalizability across different tissue regions.
An external validation dataset from The Cancer Genome Atlas containing 14 independently scanned lung adenocarcinoma cases was also used, allowing the researchers to test whether the human-in-the-loop approach transferred beyond their own institution's tissue preparation standards.
Statistical analysis used paired two-sample t-tests to compare networks with 0 versus 20 additional tumor cells, with significance set at 0.05 and a target power of 0.8.
Adding just 20 tumor cells pushed accuracy to pathologist-level performance. With 20 additional tumor cells included from the case being tested, the network achieved a mean true positive rate of 98.7 percent across all 15 in-house lung adenocarcinoma cases.
The 95 percent confidence interval for the true positive rate with 20 added cells ranged from 98.3 to 99.5 percent, which the researchers note is on par with their hypothesized 98 to 100 percent accuracy of a trained pathologist.
Across all 15 cases tested on a single high-power field, the final network showed only two false positive classifications where non-tumor cells were labeled as tumor cells. Specificity reached 99.5 percent and precision reached 99.8 percent.
Results were consistent when testing extended to three high-power fields per tumor, with the confidence interval for the true positive rate remaining at 0.983 to 0.991. Paired t-tests showed significant improvements in Dice score and true positive rate when comparing zero versus 20 added cells.
Performance on the external dataset was lower but still dramatically improved by the human-in-the-loop step. When the in-house basis network was tested on 14 lung adenocarcinoma slides from The Cancer Genome Atlas without any case-specific fine-tuning, the mean true positive rate was only 80.7 percent.
After adding 20 tumor cells from each TCGA case to the training data, the mean true positive rate jumped to 98.4 percent, a marked improvement that was statistically significant with a p-value below 0.001.
When a new basis network was trained using cells exclusively from the TCGA dataset and then tested on a held-out subset of TCGA cases, the mean true positive rate improved from 87.7 to 94.4 percent with 20 added cells, though this came with a trade-off of slightly reduced specificity.
These results demonstrate that the human-in-the-loop adaptation step is crucial when applying a network trained on one institution's tissue preparation standards to slides prepared at another institution, where staining and section thickness may differ.
The human-in-the-loop approach outperforms all previously reported general-purpose classification networks. The best network tested on the PanNuke benchmark dataset, a large heterogeneous tissue collection, achieved precision, recall, and F1-score of only 0.60, 0.71, and 0.65 for neoplastic cell classification.
The consistent performance across cases with varying tumor cell density supports the notion that the network has genuinely learned to separate lung adenocarcinoma tumor cells from the surrounding normal cell types rather than exploiting tissue density artifacts.
A comparable study classifying leukocytes in blood smears reported mean precision, recall, and F1-score of 0.9812, 0.9923, and 0.9823 respectively, suggesting that the human-in-the-loop approach generalizes well across different single-cell classification tasks in pathology.
Network training takes approximately 10 to 15 minutes, which may or may not be a practical obstacle depending on the clinical workflow. If classification results feed into downstream analyses rather than requiring the pathologist to wait at the microscope, the timing may not be a barrier to adoption.
The approach successfully reached pathologist-level classification for lung adenocarcinoma tumor cells. Across 15 different cases with five repeated tests each, the network achieved a mean true positive rate of up to 99.1 percent using only 20 case-specific cells added during microscopy review.
The output is interpretable and validatable because the network shows the pathologist which cells it classified incorrectly and with what confidence, enabling the clinician to confirm or correct predictions in real time rather than accepting opaque AI decisions.
Limitations include the homogeneity of the training data since only lung adenocarcinoma was studied, the small sample sizes in some statistical comparisons, and the need for further testing on complete whole-slide images rather than selected high-power fields.
The researchers conclude that the slide-by-slide human-in-the-loop strategy is a viable in-house solution for quantitative cell AI on standard hematoxylin and eosin stained slides, and they hope it will pave the way for broader adoption of single-cell AI tools in pathology.