A three-stage pipeline that counts and classifies cells
PlasmoCount 2.0 keeps the three-stage design of its predecessor. An object detection model first finds every red blood cell in a microscopy field of view and outputs bounding boxes. Those cells are cropped and passed to a binary classifier that decides whether each cell is infected. Infected cells then go to a regression model that predicts the parasite's developmental stage among rings, trophozoites and schizonts.
Splitting the task this way matters because infected cells are rare relative to uninfected ones. Detecting all cells first, then classifying them, avoids asking one model to handle a heavily imbalanced problem in a single step. The paper's changes sit mostly in the first two stages: a new detector, a new classifier, and a new delivery route onto the phone itself.
YOLOv8n matched larger detectors at a fraction of the size
The team fine-tuned seven detectors on their own red blood cell data: four YOLOv8 variants (nano, small, medium, large), three transformer-based models (DETR, Deformable DETR, DETA Swin Large), and the Faster R-CNN with feature pyramid network used in the original PlasmoCount. On their dataset, YOLOv8n reached an AP50 of 99.1% with 3.2 million parameters, while Faster R-CNN-FPN reached 98.5% with 41.7 million. The transformer models trailed on this dataset despite strong COCO scores: DETR 87.0, Deformable DETR 91.1, DETA Swin Large 88.7.
The size comparison is the decision-relevant part. Across the four YOLOv8 variants, AP50 stayed between 98.77% and 99.02% and the differences were not statistically significant in a paired t test. Because the larger variants bought no meaningful accuracy, the authors chose the smallest, YOLOv8n, for all further tests and deployment. The paper attributes YOLO's advantage partly to the comparatively simple morphology of red blood cells next to the complex objects in its original training data.
40x images counted cells within human disagreement
Routine microscopy is usually done at 100x with an oil-immersion lens, which ties a deep-learning tool to well-resourced labs. The authors tested whether the detector could work at 40x, where more cells fit in one image, without any 40x training data. They compared four capture strategies against human annotator counts on an independent test set of 20 images.
Plain 40x input failed: cell mean absolute error was 14.950, far worse than any annotator (2.725 to 4.550). Adding a bimodal scale augmentation during training, which spreads the scale distribution so the model sees cell sizes it would otherwise miss, brought 40x error down to 2.000 with a cell Spearman correlation of 0.954. That error sits below the spread between human annotators. The 100x strategy was still best on cell counting (MAE 1.050, correlation 0.975), and only 100x classified infected cells within the range of human error (infected MAE 0.450). The 40x bimodal and 40x cropped strategies had modestly higher infected-cell errors of 0.700 and 0.550, which the authors link to reduced resolution and less precise localisation.
Training on five species beat training on one
The training set covered 3,102 field-of-view images and 286,264 cells across five species: P. falciparum (70,277 cells), P. vivax (81,899), P. berghei (41,382), P. yoelii (48,264) and P. chabaudi (44,442). The question was whether pooling species helps or hurts when the model is tested on any single species.
It helped. An aggregate model trained on all five species outperformed species-specific training in all five species when tested on individual species, with p < 0.05 by Wilcoxon signed-rank test. The aggregate model also transferred to two species never seen in training, P. knowlesi and P. cynomolgi, evaluated on 6,494 cells from other species. The paper notes that a P. vivax-trained model appeared to do better on these unseen primate parasites than a P. falciparum-trained one, which the authors connect to P. vivax being evolutionarily closer to those species. That reading is the authors' interpretation of phylogenetic relatedness, not a tested mechanism.
ConvNeXT-tiny won on balance of accuracy and speed
For the infected-versus-uninfected classifier, the team compared several architectures against the original ResNet-50. ConvNeXT-tiny had the second-highest AUPRC at 96.43%, the best values on every other reported metric, and the second-fastest inference time. Its accuracy was 98.32%, recall 96.29%, precision 85.62%, specificity 98.55% and F1 90.60%, all as means across 5-fold cross-validation. ConvNeXT-base had the top AUPRC at 96.80% but was not chosen.
A separate addition targets whole-blood smears, where white blood cells appear at low frequency. Training the detector on red cells plus white cells raised white-cell recall at IoU 0.5 from 51.44% to 98.77% and recall across IoU 0.50-0.95 from 29.49% to 79.74%, while red-cell metrics stayed essentially flat (AP50 99.26% versus 99.25%). The white-cell evaluation used the 813 white cells in the augmented dataset, so it measures detection of those cells rather than a full diagnostic workflow.
Under three seconds on a phone, but not yet a diagnostic
The mobile app runs inference on the device and never uploads images to a server during inference. On a Google Pixel 6, a single 100x image took just under three seconds, which the paper describes as the fastest of the pipeline iterations compared. The install is 143 Mb on iOS and 206 Mb on Android, and the app supports Android 10+ and iOS 15+, so phones over a decade old can run it. Low-confidence predictions below 0.8 are flagged so a user can review them.
These are engineering results, not clinical ones. The authors state that PlasmoCount is optimised for laboratory research and that use as a clinical diagnostic aid would require sufficient regulatory approval and clinical validation. The lifecycle-stage model was left unchanged from version 1 because no new lifecycle labels were available and the existing dataset was small (n = 1,378). The paper also notes that P. malariae and P. ovale were not in the training data, so generalisation to the full set of human-infecting species remains untested.
The practical gain is throughput, not diagnosis
Read together, the results point to a specific priority. The largest measured improvements are in speed and in breadth of species coverage: a 90% reduction in processing bottlenecks, sub-three-second on-device inference, and a single model that matches or beats species-specific training across five species while transferring to two unseen ones. The 40x result matters for the same reason: it makes high-throughput counting feasible without oil-immersion microscopy, at the cost of slightly weaker infected-cell classification than 100x.
For a reader deciding whether this changes their work, the relevant split is between counting and diagnosing. If the task is parasitaemia estimation and cell counting in a research lab, the evidence supports using the aggregate model and, where throughput matters more than the last increment of accuracy, 40x imaging with scale augmentation. If the task is clinical diagnosis, the paper's own boundary applies: the tool has not been clinically validated, and the species set still excludes P. malariae and P. ovale.
Sources used in this answer
Rapid multi-species malaria parasite detection using deep learning
PlasmoCount 2.0 replaces Faster R-CNN with YOLOv8n, adds multi-species training and white-cell handling, and runs offline on a phone, processing one 100x image in under three seconds while matching or beating species-specific classifiers across five Plasmodium species.