Deep learning spots live and dead cancer cells without staining

The work, published in Applied Microscopy, addresses a familiar bottleneck in cell biology: the standard way to judge viability is to stain cells, and once a culture has been stained, its subsequent growth cannot be observed. A classifier that reads viability directly from unstained images could, in principle, allow the same dish of cells to be followed continuously in real time.

The team, led by Tomoaki Kyoden with colleagues including Shunsuke Akiguchi, Ryo Murakami, Tsugunobu Andoh and Noboru Yamada, worked with B16-BL6 cells, a mouse melanoma-derived cell line. The cells were cultured in Eagle’s minimum essential medium containing 10% fetal bovine serum and antibiotics at 37 °C in a humidified atmosphere of 5% CO2. After roughly three days of incubation the cells were transferred to 6-well culture dishes and cultured for about two more days before heating and image acquisition. The choice of an in-vitro cancer cell line matters for the study’s scope: the findings concern cultured mouse cancer cells imaged in a petri dish, not cells in a patient.

To create a sample in which living and dead cells coexisted, the researchers deliberately produced a temperature gradient across the dish. The petri dish, which had two intersecting grooves on its underside to mark the center, was placed on a support table and irradiated from above with an 808-nanometer semiconductor laser at 800 milliwatts for 2 minutes 30 seconds. A graphite cylinder under the dish made contact at the center, and thermocouples with a wire diameter of 0.1 millimeter, positioned 2.0, 5.0 and 10 millimeters from the center, recorded the underside temperature every 0.5 seconds on a data logger. Spot heating the center killed cells near the middle, while the temperature fell toward room temperature with distance from the center, so the proportion of surviving cells rose toward the periphery.

Imaging was done with an inverted phase-contrast microscope, model XD30-PH, with a camera connected through a C-mount. Using the cross-shaped grooves to locate the center, the researchers moved the dish outward in 2-millimeter increments with a micrometer stage and captured eight observation images of 1920 by 1080 pixels. Each position was photographed at three stages: before laser heating, after laser heating but before staining, and after staining. Accounting for the objective lens magnification of 4×, a C-mount magnification of 1×, a 13-inch monitor, a 6-millimeter camera sensor diagonal and a zoom factor of 1×, the effective magnification was calculated as 220×.

Staining was performed with trypan blue, a dye that enters cells whose membranes have been destroyed and is widely used to judge viability in petri dishes. After heating, 5 milliliters of a commercially available trypan blue solution was spread across the bottom of the dish and then removed before the stained images were taken. From the stained images, the team selected regions where visual inspection clearly showed live or dead cells, increased the saturation, and manually drew mask images with gray for live regions, white for dead regions and black for other, non-cell regions. These masks served as ground truth labels for supervised training.

The images were cut into segments of 60 by 60 pixels, a size chosen so that each segment approximately contained a single cell. Twenty-four phase-contrast microscope images yielded 576 segments each, for a total of 13,824 segmented images. Of these, 11,059, or 80%, were used for training and 2,765 for validation. All 11,059 training images were augmented by three 90-degree rotations, producing 44,236 training images, and one training epoch comprised those 44,236 images. Training ran for up to 99 epochs. On stained validation images, the percentage of correct responses reached 0.975 at epoch 99, and the program correctly discriminated dead cells from live cells even in complicated images where the two were intermixed, a task the authors note is difficult by visual inspection alone.

The key step came next. Because the pre-heating, post-heating and post-staining images were taken at the same positions, the classification results from the stained images could be overlaid onto the corresponding unstained images. Wherever the stained-image classification at a position matched, the unstained image segment could be labeled live or dead, generating training and validation data for a second program that classifies viability from unstained images alone. That second program achieved an accuracy of 0.931 after 99 epochs, which the authors describe as performance comparable to conventional stained-image analysis. The approach also enabled estimation of the spatial boundary between live and dead cell populations across the dish.

Agreement between the stained and unstained classifiers was not uniform across the dish. For each observation image taken at 2-millimeter intervals from the center toward the edge, the researchers calculated the ratio of segments classified the same way before and after staining. Except at one position, the agreement rate was approximately 90% everywhere, close to the correct response rate on stained images. At 4 millimeters from the center, however, agreement dropped to about 68.9%. This was the region where viable and dead cells were mixed and where some cells could die without noticeable changes from their original shape. There, cells that trypan blue showed to be viable were often classified as dead in the unstained images. At 6 millimeters, where accuracy was relatively high, the match rate was 94.3%.

Despite that local weakness, the method captured the overall geography of cell death. When live and dead cells are present in a dish and survival varies spatially in a continuous manner, the technique can identify the point at which the ratio of live to dead cells reverses, the 50% death rate. In the images reported, that 50% position coincided within a difference of 0.252 millimeters between the stained and unstained analyses. The authors suggest this means the boundary location can be recovered even where slight morphological changes are misleading, by comparing stained and unstained results and extrapolating the viability distribution.

The researchers point toward drug evaluation as a possible extension, noting that drug-induced cell death assessment is common in cell and pharmacological research and that anticancer drugs can induce characteristic morphological changes, with databases such as Morphobase proposed to categorize them. They are careful to add that morphological alterations do not always directly correspond to cell death, and that further validation remains necessary before the label-free method could enter drug-evaluation workflows. Within its demonstrated scope, though, the result stands: a program trained on stained images can teach a second program to read viability from unstained phase-contrast images of cultured cancer cells with 0.931 accuracy, opening the possibility of observing how cells respond over time without ever staining them.

Subject of Research: Deep learning classification of live and dead unstained cancer cells in microscopic images

Article Title: Viability classification of unstained cells in microscopic images using deep learning

Article References: Kyoden, T., Akiguchi, S., Murakami, R., Andoh, T., & Yamada, N. (2026). Viability classification of unstained cells in microscopic images using deep learning. Applied Microscopy, 56(1), Article 7. https://doi.org/10.1186/s42649-026-00127-9

Image Credits: AI Generated

DOI: 10.1186/s42649-026-00127-9

Keywords: deep learning, cell viability, phase-contrast microscopy, trypan blue staining, cancer cells, B16-BL6 melanoma, unstained images, laser heating, image classification, in vitro culture, supervised training, label-free imaging