Organizations: Department of Physics, University of Geneva, Geneva, Switzerland · Geneva School of Engineering, Architecture and Landscape (HEPIA), HES-SO University of Applied Sciences and Arts Western Switzerland, Geneva, Switzerland · School of Engineering and Architecture of Fribourg (HEIA-FR), HES-SO University of Applied Sciences and Arts Western Switzerland, Fribourg, Switzerland · Axon Logic, Multidisciplinary Research and Innovation Centre, Athens, Greece
Limited annotated data, class imbalance, and morphological similarity make white blood cell classification challenging. We benchmark hybrid variational quantum classifier heads against an affine baseline on 821 microscopy images across seven morphological classes. Using ImageNet-initialised ResNet34 backbones and repeated stratified five-fold cross-validation, we first compare direct transfer with intermediate adaptation on MLL23, a larger peripheral blood cell microscopy dataset. To guide the head comparison, we explore different encoding strategies, variational depth, and the number of independent circuit blocks. We then compare the classifier performance using the affine head and two hybrid heads across multiple random seeds, under a shared training and evaluation protocol. For this comparison, all backbones share the same MLL23-adapted initialisation. Intermediate adaptation increases mean macro-F1 across all four affine-head fine-tuning configurations. The largest gain occurs with a frozen backbone, where macro-F1 rises from 0.685 to 0.786, a relative improvement of approximately 14.7%. The shallow two-block hybrid head achieves a mean pooled out-of-fold macro-F1 of 0.865±0.010, compared with 0.879±0.008 for the affine head. The corresponding balanced accuracies are 0.886±0.008 and 0.889±0.010, respectively. In the architectural exploration, the two-block hybrid head achieves the highest macro-F1 and balanced accuracy among the evaluated hybrid configurations, while the effect of encoding varies across the considered metrics. Our results help guide the design and evaluation of hybrid classifier heads for imbalanced microscopy by linking architectural choices to overall performance and class-specific precision-recall trade-offs.
Figures & tables
Class label
Cell type
Samples
BAS
Basophil
18
EOS
Eosinophil
43
KSC
Smudge cell
158
LYA
Atypical lymphocyte
30
LYT
Typical lymphocyte
224
MON
Monocyte
180
Table 1: Class distribution of the final MicroscopyRGB dataset ( N=821 ).
Figure 1: Affine and hybrid classification pipelines used for the head comparison. Both ResNet34 backbones start from the same MLL23-adapted checkpoint and are fine-tuned separately. The hybrid head uses B independent 12-qubit blocks, each with L variational layers. The projection dimension is D=36B for standard re-uploading and D=36BL for incremental re-uploading. Concatenation yields 36B expectation values for the classical readout. Input preprocessing and the tanh bounding in Eq. 3 are omitted
Figure 2: Structure of one 12-qubit VQC block b . (a) Each of the L variational layers applies RX , RZ , and RY rotations sequentially, with trainable angle offsets and encoding scales, followed by a trainable Ising- ZZ ring. The final Roti=Rot(αi,βi,δi) gates are applied once, after layer L . Expectation values of X , Y , and Z on each qubit form a 36-component readout vector. (b) Nearest-neighbour ring topology, including the 12-1 edge. The block index b is suppressed in local gate labels
Model
Max. circuit width
Post-backbone head
Circuit params.
Total
Trainable
ResNet34 + affine head
-
3,591
-
21,288,263
19,940,359
Hybrid B1/L1
12
19,943
120
21,304,615
19,956,711
Hybrid B2/L1
12
38,855
240
21,323,527
19,975,623
Hybrid B3/L1
12
57,767
360
21,342,439
19,994,535
Table 2: Parameter counts for the shallow MLL23-adapted configurations. Hybrid rows use L=1 , an affine angle projection with tanh bounding, Bloch-vector readout, and trainable Ising- ZZ and final measurement-basis rotations. Trainable totals assume the layer-3-4 fine-tuning configuration. Circuit parameters are included in the post-backbone head count.
Figure 3: Macro-F1 scores and confusion matrices with and without intermediate MLL23 adaptation. (a) Mean pooled out-of-fold (OOF) macro-F1 for four ResNet34 fine-tuning configurations. Error bars show sample standard deviations across seeds 9, 122, and 250. (b) Confusion matrices for frozen-backbone models using direct ImageNet transfer and MLL23-adapted initialisation. For each seed, the pooled OOF matrix is normalised by true class, and corresponding entries are then averaged across the three seeds. Cell values and colours indicate mean within-class percentages. Diagonal entries represent mean class recall. All configurations use affine classifier heads and five-fold cross-validation on the same 821 images for each seed
Figure 4: Comparison of the affine head and the two selected hybrid heads. (a) Pooled OOF macro-F1. (b) Balanced accuracy. (c) Class-specific precision, recall, and F1 for BAS ( n=18 ), LYA ( n=30 ), and EOS ( n=43 ). Bar labels show means across seeds 9, 122, and 250, and error bars show standard deviations. Metrics are calculated from pooled OOF predictions separately for each seed. All models start from the same MLL23-adapted ResNet34 weights and fine-tune residual layers 3-4 jointly with the classifier head
Encoding
Config.
Aggregate metrics
BAS
LYA
Macro-F1
Balanced accuracy
Precision
Recall
Precision
Recall
Single encoding
B1/L1
0.866
0.873
0.824
0.778
0.415
0.567
B2/L1
0.869
0.888
0.750
0.833
0.435
0.667
B3/L1
0.854
0.865
0.750
0.833
0.385
0.500
Standard re-uploading
B1/L2
0.804
0.858
0.452
0.778
0.292
0.700
Incremental re-uploading
B1/L2
0.830
0.847
0.667
0.667
0.327
0.600
Table 3: Macro-F1, balanced accuracy, and class-specific precision and recall for seven MLL23-adapted hybrid configurations evaluated using a single seed.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Backbone fine-tuning
Initialisation
Accuracy
Balanced Accuracy
Macro-Precision
Macro-F1
Weighted-F1
Full
ImageNet
0.935±0.004
0.891±0.009
0.861±0.008
0.874±0.008
0.937±0.003
Full
MLL23-adapted
0.933±0.004
0.901±0.012
0.869±0.009
0.881±0.004
0.936±0.001
Layers 3-4
ImageNet
0.932±0.004
0.895±0.004
0.862±0.004
0.875±0.004
0.935±0.002
Layers 3-4
MLL23-adapted
0.935±0.009
0.889±0.010
0.872±0.008
0.879±0.008
0.936±0.008
Layer 4 only
ImageNet
0.921±0.003
0.866±0.020
0.835±0.009
0.847±0.009
0.925±0.003
Layer 4 only
MLL23-adapted
0.929±0.003
0.869±0.007
0.849±0.003
0.858±0.002
0.930±0.003
Appendix
Table S1: Transfer-learning comparison using ResNet34 models with affine classifier heads. Values are the mean ± sample standard deviation of pooled OOF metrics across three seeds.
Class
Support
Initialisation
Precision
Recall
F1
BAS
18
ImageNet
0.229±0.011
0.463±0.032
0.307±0.015
MLL23-adapted
0.408±0.030
0.741±0.085
0.526±0.040
EOS
43
ImageNet
0.819±0.013
0.806±0.013
0.812±0.012
MLL23-adapted
0.939±0.012
0.829±0.027
0.881±0.010
KSC
158
ImageNet
0.941±0.034
0.916±0.013
0.928±0.010
MLL23-adapted
0.976±0.004
0.960±0.013
0.968±0.009
Appendix
Table S2: Class-specific performance with and without intermediate MLL23 adaptation for ResNet34 models with a frozen backbone and an affine classifier head.
Head
Accuracy
Balanced Accuracy
Macro-Precision
Macro-F1
Weighted-F1
Affine
0.935±0.009
0.889±0.010
0.872±0.008
0.879±0.008
0.936±0.008
Hybrid B1/L1
0.918±0.011
0.876±0.008
0.842±0.020
0.854±0.013
0.922±0.010
Hybrid B2/L1
0.924±0.009
0.886±0.008
0.850±0.010
0.865±0.010
0.928±0.008
Appendix
Table S3: Aggregate affine-hybrid comparison using the MLL23-adapted ResNet34 backbone with residual layers 3-4 fine-tuned. In the configuration labels, B denotes the number of independent 12-qubit blocks and L the number of variational layers per block. Values are the mean ± sample standard deviation of pooled OOF metrics across three seeds.
Class
Head
Precision
Recall
F1
BAS
Affine
0.798±0.041
0.870±0.032
0.832±0.029
Hybrid B1/L1
0.710±0.103
0.852±0.085
0.768±0.047
Hybrid B2/L1
0.715±0.043
0.870±0.032
0.784±0.020
LYA
Affine
0.478±0.076
0.567±0.058
0.517±0.063
Hybrid B1/L1
0.385±0.070
0.567±0.100
0.458±0.081
Hybrid B2/L1
0.423±0.070
0.589±0.069
0.490±0.060
Appendix
Table S4: Pooled OOF precision, recall, and F1 for the three least-supported classes: basophils (BAS, n=18 ), atypical lymphocytes (LYA, n=30 ), and eosinophils (EOS, n=43 ). All pipelines use the MLL23-adapted ResNet34 backbone with residual layers 3-4 fine-tuned. Both hybrid heads use L=1 , with independently simulated 12-qubit blocks. Values are the mean ± sample standard deviation across three seeds.
Configuration
Trainable head params.
CV runtime (h)
Affine
3591
0.224
Single encoding B1/L1
19943
22.716
Single encoding B2/L1
38855
44.333
Single encoding B3/L1
57767
66.698
Standard re-uploading B1/L2
20027
34.613
Incremental re-uploading B1/L2
38495
34.519
Appendix
Table S5: Observed end-to-end training runtimes and trainable head parameter counts for the affine baseline and seven hybrid configurations. Each runtime is the total time recorded in Weights & Biases across the five cross-validation folds for seed 250 . All configurations are trained on a single NVIDIA A40 GPU for 50 epochs per fold. Hybrid circuits are evaluated by noiseless statevector simulation.
Figure S1: Pooled OOF confusion matrices for seed 9. From left to right: Affine, Hybrid B1/L1 , and Hybrid B2/L1 . Cell annotations indicate image counts, whereas colour represents percentages within each true class
Figure S2: Pooled OOF confusion matrices for seed 122. From left to right: Affine, Hybrid B1/L1 , and Hybrid B2/L1 . Cell annotations indicate image counts, whereas colour represents percentages within each true class
Figure S3: Pooled OOF confusion matrices for seed 250. From left to right: Affine, Hybrid B1/L1 , and Hybrid B2/L1 . Cell annotations indicate image counts, whereas colour represents percentages within each true class
Accurate classification of microscopic blood cells is still a critical task in medical image analysis, where subtle variations and limited data can challenge conventional deep learning models. As such, we investigate in this work the potential of Hybrid Quantum-Classical Neural Networks (HQNNs) to enhance feature representation and improve classification performance in this domain. We propose a modular architecture combining a pre-trained ResNet-50 backbone with a low-dimensional latent bottleneck and a variational quantum circuit, enabling a direct comparison between quantum-enhanced and purely classical transformation mechanisms. To isolate the contribution of the quantum component, we evaluate three architectures: a HQNN model, a Classical Matched Model with an additional nonlinear transformation layer of comparable capacity, and a baseline model without an intermediate transformation stage. Experiments conducted on two publicly available blood cell datasets, namely the Blood Cell Images dataset and the PBC dataset, demonstrate that HQNNs consistently achieve superior or more balanced performance across evaluation metrics. In the Blood Cell Images Dataset, the proposed approach improves macro F1-score by up to 3.7% compared to classical baselines, while improving the F1-score from 98.54% to 98.69% in the more challenging 8-class scenario with near-saturated performance. Additional evaluation on IBM quantum hardware shows that the model remains robust under noise, with only a modest performance degradation relative to simulated results. These results indicate that quantum feature transformations can enhance discriminative representations, particularly in challenging classification scenarios, and highlight the practical potential of HQNN models for medical imaging tasks.
Guilherme Cruz, Nouhaila Innan, Alberto Marchisio +2
Instituto de Telecomunicac¸˜oes, University of Coimbra, Portugal · eBrain Lab, Division of Engineering, New York University Abu Dhabi, PO Box 129188, Abu Dhabi, UAE · Center for Quantum and Topological Systems, NYUAD Research Institute, New York University Abu Dhabi, UAE +1
Automated white blood cell (WBC) classification is essential for leukemia screening but remains challenged by extreme class imbalance, long-tail distributions, and domain shift, leading deep models to overfit dominant classes and fail on rare subtypes. We propose a hybrid framework for rare-class generalization that integrates a generative Pix2Pix-based restoration module for artifact removal, a Swin Transformer ensemble with MedSigLIP contrastive embeddings for robust representation learning, and a biologically-inspired refinement step using geometric spikiness and Mahalanobis-based morphological constraints to recover out-of-distribution predictions. Evaluated on the WBCBench 2026 challenge, our method achieves a Macro-F1 of 0.77139 on the private leaderboard, demonstrating strong performance under severe imbalance and highlighting the value of incorporating biological priors into deep learning for hematological image analysis. The code is available at https://github.com/trongduc-nguyen/WBCBench2026
Duc T. Nguyen, Hoang-Long Nguyen, Huy-Hieu Pham
College of Engineering & Computer Science, VinUniversity, Hanoi, Vietnam · VinUni-Illinois Smart Health Center, VinUniversity, Hanoi, Vietnam · The Computer Vision and Medical AI Lab, VinUniversity, Hanoi, Vietnam
Automating white blood cell classification for diagnosis of leukaemia is a promising alternative to time-consuming and resource-intensive examination of cells by expert pathologists. However, designing robust algorithms for classification of rare cell types remains challenging due to variations in staining, scanning and inter-patient heterogeneity. We propose a lightweight ensemble approach for classification of cells during Haematopoiesis, with a focus on the biology of Granulopoiesis, Monocytopoiesis and Lymphopoiesis. Through dataset expansion to alleviate some class imbalance, we demonstrate that a simple ensemble of lightweight pretrained SwinV2-Tiny, DinoBloom-Small and ConvNeXT-V2-Tiny models achieves excellent performance on this challenging dataset. We train 3 instantiations of each architecture in a stratified 3-fold cross-validation framework; for an input image, we forward-pass through all 9 models and aggregate through logit averaging. We further reason on the weaknesses of our model in confusing similar-looking myelocytes in granulopoiesis and lymphocytes in lymphopoiesis. Code: https://gitlab.com/siddharthsrivastava/wbc-bench-2026.
Siddharth Srivastava, Adam Smith, Scott Brooks +2
Department of Computer Science, University of Warwick · Warwick Medical School, University of Warwick · Intelligent Imaging Innovations Ltd