Ki-67 is a protein found on the surface of dividing cancer cells. Pathologists use it as a proliferation marker -- the more Ki-67-positive cells in a tumor, the faster it is growing. A high Ki-67 score can indicate a poor prognosis and influence treatment choices.
To detect Ki-67, labs currently use a second staining procedure called immunohistochemistry (IHC), which applies targeted antibodies to highlight specific proteins. This step adds cost, time, and the need for extra tissue, making it unavailable for every patient.
The most common stain in pathology is hematoxylin and eosin (H&E), which colors cell structures in shades of blue and pink and is routinely available for virtually all cancer biopsies. H&E shows cell shape (morphology) but has traditionally been thought to reveal nothing about molecular markers like Ki-67.
This study challenges that assumption. The authors hypothesized that if Ki-67 positivity causes subtle changes in how cells look, a deep learning model trained on H&E images could learn those visual cues and predict Ki-67 status without any additional IHC staining.
Tissue samples from 12 patients with neuroendocrine tumors of the gastrointestinal tract were used. From each tumor block, two consecutive thin slices were cut -- one stained with H&E and one with the Ki-67 antibody -- so corresponding regions could be directly compared.
From the H&E slides, researchers identified 300 regions of interest (ROIs) and manually annotated each visible cell as either Ki-67 positive or Ki-67 negative, using the paired IHC slide as the ground truth reference. This produced over 18,700 individual cell image patches sized 64 x 64 pixels.
The dataset included roughly equal numbers of positive cell patches, negative cell patches, and background patches (areas containing no cells). Keeping background as a third class helps the model learn to ignore non-cellular tissue regions.
The full dataset was split 80% for training and 20% for validation, and a 10-fold cross-validation strategy was also applied to confirm that results were stable regardless of how the data were divided.
The study used a modified version of ResNet18, a well-established deep convolutional neural network architecture. ResNet (Residual Network) uses shortcut connections between layers that make very deep networks easier to train without losing accuracy.
The original ResNet18 was adapted for this task by removing the final average pooling layer and changing the output to three classes (Ki-67 positive, negative, background). The model was pre-trained on ImageNet -- a massive database of natural photographs -- before being fine-tuned on the histopathology data, which accelerated training.
Training used cross-entropy loss and the Adam optimizer over 20 epochs. Data augmentation techniques such as random horizontal and vertical flipping were applied to reduce overfitting and improve the model's ability to generalize to new slides.
To apply the classifier across entire large-scale ROIs (not just 64x64 patches), the authors converted all fully connected layers into convolutional layers, creating what is known as a fully convolutional network (FCN). This transformation allows the model to process images of any size and output a spatial classification map showing the predicted location of every positive and negative cell.
The model achieved 93.71% average accuracy on the validation set, correctly identifying Ki-67 positive cells, negative cells, and background patches. On the training set, accuracy reached 97.80%, and 10-fold cross-validation confirmed a mean accuracy of 93.10%, showing the results were consistent and robust.
Performance was especially high for background classification (98.5% precision), which makes sense since empty tissue is visually distinct from cells. Positive and negative cells were harder to distinguish, with precision around 91-91%, reflecting the subtle visual differences between them in H&E staining.
For quantification, the model's predictions across 32 large ROIs were compared to actual IHC measurements. The correlation coefficient for the Ki-67 positive rate (Rpos) between model-predicted H&E maps and real IHC stains was 0.80, which indicates a strong and statistically meaningful agreement.
Cases with moderate and low Ki-67 positivity correlated better than high-density positive cases. In tumors densely packed with positive cells, the irregular mixing of positive and negative cells made it harder to annotate training data accurately and therefore more difficult for the model to learn the distinction.
The most important conceptual finding is that molecular information is encoded in cell morphology. The fact that a model trained only on H&E images can predict Ki-67 status supports the long-held hypothesis that the microscopic appearance of cells reflects the underlying molecular processes driving their behavior.
This aligns with earlier research showing that deep learning can predict genetic mutations and other molecular features from H&E pathology slides, suggesting that routine staining carries far more information than pathologists have traditionally been able to extract visually.
The authors candidly identify current limitations. Model accuracy depends heavily on image quality, which in turn depends on slide preparation, stain consistency, and scanner precision. Variability in any of these steps could reduce reliability in real-world clinical settings.
The relationship between morphology and molecular markers is also likely to be complex and tissue-specific. This study focused on a single marker (Ki-67) in a single tumor type (neuroendocrine tumors). Future work will need to validate the approach across diverse cancers, tissues, and antibody targets before it can be broadly applied clinically.
In current clinical practice, a patient who has already had an H&E biopsy often needs a separate IHC test to determine their tumor proliferation rate, which informs treatment decisions. This second test requires additional laboratory time and cost, and is not always available in lower-resource settings.
If a deep learning model can reliably predict Ki-67 from existing H&E slides, it could allow physicians to access this prognostic information without additional tissue sampling or staining, accelerating diagnostic turnaround and reducing healthcare costs.
This is especially relevant for neuroendocrine tumors, where Ki-67 index is used to grade the tumor (G1, G2, or G3) and directly guides whether surgery, chemotherapy, or targeted therapy is recommended. Faster Ki-67 quantification could speed treatment planning for these patients.
The broader vision is a future where a single routine H&E scan provides not just morphological grading but also a panel of molecular biomarker predictions, essentially transforming one routine test into many using AI. The authors plan to expand the dataset and test additional cancer types and molecular markers in future work.
This study demonstrates that a modified ResNet18 deep learning model can predict Ki-67 positive cells directly from H&E stained slides with high accuracy, achieving 93.71% classification accuracy and a 0.80 correlation with real IHC measurements.
The results provide evidence that molecular markers are not invisible in H&E images -- they leave detectable morphological signatures that deep neural networks can learn to recognize, even when the signals are too subtle for pathologists to perceive.
Future plans include enlarging the dataset, testing different cancer and tissue types, and optimizing the model with semi-supervised learning to reduce the burden of manual annotation. The authors also plan to investigate whether similar methods can predict other clinically important molecular markers beyond Ki-67.
The code for this project is publicly available, facilitating reproducibility and enabling other researchers to build on this approach for different biomarkers and tumor types.