cAItomorph: Transformer AI for Blood Malignancy

• cAItomorph is an explainable, transformer-based AI system that predicts hematological malignancies from peripheral blood smear images using weak supervision.
• It employs multiple-instance learning with transformer aggregation and incorporates multimodal data like age, sex, and blood counts to enhance diagnostic accuracy.
• The system reduces unnecessary invasive bone marrow procedures by providing reliable, interpretable patient-level predictions via attention mapping.

cAItomorph is an explainable, transformer-based weakly supervised AI system for predicting hematological malignancies from peripheral blood smear single-cell images. Its clinical aim is to use the peripheral blood smear as a rapid, minimally invasive screening and triage modality for diseases whose definitive diagnosis usually requires bone marrow aspiration, while preserving patient-level diagnostic grounding through bone-marrow-based cytomorphology, cytogenetics, molecular genetics, and immunophenotyping. In the reported real-world cohort, the system was trained and evaluated on peripheral blood single-cell images from 6115 patients and 495 healthy controls, with predictions organized into 22 detailed and 7 coarse diagnostic classes (Dasdelen et al., 23 Sep 2025).

1. Clinical setting and cohort design

cAItomorph was developed for a setting in which peripheral blood smears remain a cornerstone in the diagnosis of hematological neoplasms, yet neoplastic transformations typically arise in the bone marrow and may not manifest as detectable aberrations in peripheral blood. This creates a diagnostic problem: blood smears are cheap and rapid, but they do not always permit confident confirmation or exclusion of malignancy, so invasive follow-up remains common.

The underlying cohort is explicitly described as a real-world cohort from the Munich Leukemia Laboratory. It includes 6115 patients with first-time diagnoses from 2021–2022 and 495 healthy donors. All patient diagnoses were established using bone-marrow-based ground truth from cytomorphology, immunophenotyping, cytogenetics, and molecular genetics, following WHO guidelines. Each patient has roughly $\sim 500$ single white blood cell images from peripheral blood smears, stored as $144\times 144$ TIFFs, and the full dataset contains over 3.2 million single-cell images. The data were also augmented with age and sex, and in the multimodal variant with blood count information (Dasdelen et al., 23 Sep 2025).

This cohort construction is central to the system’s intended use. Rather than relying on a curated benchmark, the model is trained on heterogeneous routine-practice material, so its reported behavior is tied to the diagnostic distribution and uncertainty structure of real hematology workflows.

2. Diagnostic taxonomy and dataset curation

The raw cohort contains 168 diagnostic labels. These were grouped into 22 detailed classes and, for the main task, into 7 coarse classes: acute leukemia, lymphoma, MDS, MDS/MPN, MPN, plasma cell neoplasm, and no malignancy. The coarse-class composition reflects the real cohort distribution: acute leukemia is mostly AML with a smaller ALL and “other AL” component; lymphoma is dominated by B-cell neoplasms and CLL; MPN includes CML, ET, PV, and many “other MPN” cases; and no malignancy includes healthy donors and reactive changes.

Model development used a cleaned subset obtained by excluding post-chemotherapy follow-ups (3165), unclear diagnoses (95), double / in-between diagnoses (1148), and rare or undetectable conditions (143). After this filtering, the main dataset used for model development contained 2059 patients and 1,003,702 single-cell images (Dasdelen et al., 23 Sep 2025).

This curation defines the article’s central distinction between diagnostic scope and diagnostic ambiguity. The 7-class problem is not a generic image-classification benchmark; it is a patient-level differential diagnosis problem constructed from clinically heterogeneous labels, with exclusions designed to stabilize the main cross-validation experiments. A plausible implication is that the cleaned training cohort is optimized for reproducible model development, while the extended test analyses probe behavior closer to the edges of routine diagnostic practice.

3. Architecture and weakly supervised aggregation

cAItomorph follows a weakly supervised multiple-instance learning style pipeline in which single-cell images are encoded first, then aggregated into a patient representation, and finally classified. The best-performing model uses DinoBloom, a hematology foundation model trained self-supervised on 13 public white blood cell datasets and more than 380,000 single-cell images. DinoBloom is used as a frozen encoder that maps each $144\times 144$ cell image into a 768-dimensional embedding.

For each patient, around 500 cell embeddings are passed as tokens into a transformer aggregator. The best transformer aggregator has 2 layers and 8 attention heads per layer. It produces a class token that serves as the patient-level representation. This replaces a purely MIL-style independent weighting scheme with a mechanism that can model interactions among cells within the same patient. The transformer output can then be combined with age, sex, and optionally blood counts; these metadata are normalized, passed through a linear layer with ReLU and layer norm, and concatenated with the image-derived representation. The final classifier is a small MLP with two fully connected layers, 64 neurons, and ReLU between them (Dasdelen et al., 23 Sep 2025).

The paper compares five architectures and reports a monotonic improvement as encoder and aggregator sophistication increase:

Model	Architecture	Mean accuracy
1	ResNet34 + MIL attention aggregator	61.3%
2	ViT + MIL aggregator	64.2%
3	ViT + transformer aggregator	65.9%
4	DinoBloom + transformer aggregator	66.7%
5	DinoBloom + transformer aggregator + age/sex	67.6%

Training is performed with the encoder frozen, cross-entropy loss, AdamW, a learning rate of $2\times 10^{-5}$, 40 epochs, and learning-rate scheduling with ReduceLROnPlateau; model selection uses best loss. Evaluation uses 5-fold cross-validation with patient-level splitting, so each patient appears in the test set exactly once (Dasdelen et al., 23 Sep 2025).

4. Mathematical formulation, calibration, and predictive performance

The supplementary text gives the classical MIL-style attention formulation

$z_c = \sum_{j=1}^{N} \alpha_{c,j} h_j,$

with

$$\alpha_{c,j} = \frac{\exp\left[w^\top \tanh(V h_j)\right]} {\sum_{l=1}^{N}\exp\left[w^\top \tanh(V h_l)\right]},$$

where $h_j$ is the embedding of the $j$-th cell image and $V \in \mathbb{R}^{D \times L}$, $w \in \mathbb{R}^{D}$, with $144\times 144$0, are trainable parameters. Although the strongest model uses transformer aggregation, this formulation clarifies the paper’s broader multiple-instance learning frame and its shift from per-cell attention weights toward contextual token processing.

On the cleaned 7-class task, the best model achieves 67.6 $144\times 144$1 0.8% overall accuracy; the abstract rounds this to 68 $144\times 144$2 1%. The reported F1 scores are 74 $144\times 144$3 2% for acute leukemia, 75 $144\times 144$4 3% for MPN, and 82 $144\times 144$5 3% for no malignancy. Coarse-class sensitivities are 77.0 $144\times 144$6 5.0% for acute leukemia, 42.6 $144\times 144$7 5.8% for lymphoma, 46.4 $144\times 144$8 7.9% for MDS, 37.2 $144\times 144$9 9.7% for MDS/MPN, 78.5 $144\times 144$0 6.4% for MPN, 58.7 $144\times 144$1 6.0% for plasma cell neoplasm, and 85.1 $144\times 144$2 3.1% for no malignancy. For the binary malignant-vs-healthy split, the model reaches 92.7 $144\times 144$3 0.7% precision and 93.3 $144\times 144$4 0.9% sensitivity (Dasdelen et al., 23 Sep 2025).

Top-2 prediction analysis increases overall accuracy to 84.3 $144\times 144$5 1.3%. This is especially notable for more difficult categories, where the correct diagnosis is often retained among the strongest differentials: 69.9% for lymphoma, 75.8% for MDS, 73.5% for MDS/MPN, and 81.2% for plasma cell neoplasm. The paper also emphasizes calibration, using reliability diagrams and Expected Calibration Error, and reports an ECE of 5.35 while describing the model as inherently well-calibrated. This calibration layer is important because the system is intended not only to rank diagnostic possibilities but also to emit clinically interpretable prediction probabilities (Dasdelen et al., 23 Sep 2025).

5. Explainability and diagnostic triage

Explainability is treated as a core property rather than an auxiliary visualization. For patient-level attribution, the transformer model uses Attention Rollout to rank which single-cell images influenced the final diagnosis most. For selected high-attention cells, Score-CAM produces pixel-level heatmaps that identify image regions associated with the prediction. The paper gives clinically specific examples: myeloblasts in AML, giant thrombocytes in MPN and PV, neoplastic or reactive lymphocytes in lymphoma-like cases, overcrowded erythrocytes and thrombocytes in PV, and myeloblast nuclei with surrounding cytoplasm in AML and MDS/AML borderline cases.

The model also visualizes attention distribution over multiple heads, supporting the claim that the transformer learns different contextual relationships among cells. This interpretability framework is presented as a means for hematologists to inspect the most relevant cells first, understand why a diagnosis is suggested, assess confidence in borderline cases, and identify cases needing further workup (Dasdelen et al., 23 Sep 2025).

A distinct clinical application is bone marrow aspiration recommendation. For each disease class, the model outputs a probability $144\times 144$6; if this exceeds a class threshold $144\times 144$7, the system recommends aspiration. A hyperparameter $144\times 144$8 tunes sensitivity, and acute leukemias are treated more conservatively by making the model five times more sensitive for acute leukemia. At the selected operating point, the baseline clinic false discovery rate is 13.8%, whereas cAItomorph reduces it to 12.0%, a 13% relative reduction in unnecessary bone marrow aspirations. Crucially, no acute leukemia cases were missed, and other disease categories had less than 2% oversight (Dasdelen et al., 23 Sep 2025).

6. Strengths, edge cases, and clinical significance

At the 22-class level, performance is strongest for morphologically obvious or high-burden diseases. Reported accuracies include 76.6 $144\times 144$9 4.1% for AML, 80.1 $2\times 10^{-5}$0 13.3% for ALL, 86.4 $2\times 10^{-5}$1 12.9% for CLL, 100 $2\times 10^{-5}$2 0% for ET, 80.0 $2\times 10^{-5}$3 40.0% for PV, 82.8 $2\times 10^{-5}$4 12.6% for CML, and 98.8 $2\times 10^{-5}$5 0.7% for healthy donors. Performance is lower for entities that are hard to diagnose from peripheral blood alone, including hairy cell leukemia, some lymphoma subtypes, MDS, multiple myeloma, and reactive changes. The paper specifically notes that “other MDS/MPN” cases are often classified as MPN, which it describes as clinically plausible given overlapping morphology.

The model was also tested on an extended test set that includes unclear diagnoses, double / in-between diagnoses, MGUS-related cases, and borderline MDS/AML and MDS/MPN cases. In that setting, 72.7% of MDS/AML borderline cases are classified as acute leukemia and 37.0% of MGUS cases are classified as plasma cell neoplasm. On a public external AML dataset, the model reports 96% sensitivity for AML detection, though some healthy donors were misclassified because of domain shift (Dasdelen et al., 23 Sep 2025).

These results delimit both the promise and the constraints of the system. cAItomorph is strongest where peripheral blood cytomorphology carries informative malignant signal, and weaker where peripheral abnormalities are absent, subtle, or diagnostically overlapping. The broader significance claimed for the model is therefore not replacement of bone-marrow-based diagnosis, but AI-assisted hematology diagnostics: a foundation-model-based, weakly supervised, transformer-aggregated framework that can support differential diagnosis, improve triage, reduce unnecessary invasive procedures, and do so on heterogeneous real-world data. The paper also states that code, test data, and model weights are publicly available to ensure reproducibility (Dasdelen et al., 23 Sep 2025).