Sofia Vargas Ibarra

Thrombus imaging with AI for patients stratification S3Seminar Sof´ ıa
Vargas Ibarra Postdoctoral Researcher at L2S Universit´ e d’´ Evry, Centre Hospitalier Sud Francilien October 17, 2025

Thrombus imaging with AI for patients stratification Stroke patient care
1 / 30

Thrombus imaging with AI for patients stratification Stroke 1st cause
of death in women 1 over 6 people 24 millions per year 1st cause of handicap 2 / 30

Thrombus imaging with AI for patients stratification Time is brain:
Every second, 1.9M dead neurons 3 / 30

Thrombus imaging with AI for patients stratification Distal vs Proximal
4 / 30

Thrombus imaging with AI for patients stratification Treatment and outcomes:
Thrombolysis and Thrombectomy Fig 1: Recanalization MRIt0-MRIt1 → 1/0: recovered/another treatment 5 / 30

Thrombus imaging with AI for patients stratification State of the
art State of the art Multicentric Robustness for missing modalities: syntatic modality generation, training strategies or specific architectures. Automatic ROI Characterization: classical image processing methods and deep learning segmentation architectures. Stroke evolution assesment: automatic software and biomarkers Limitations ◦ Focused on CT scans (no distal thrombi visible) ◦ No relationship between lesion and thrombus ◦ High training and optimization complexity 6 / 30

Thrombus imaging with AI for patients stratification State of the
art Goal: Predict probability of recanalization success For stroke patients with distal thrombus 1h after thrombolysis Automatic tool for segmenting lesion and thrombus using MRI, with multicentric robustness, including missing modalities scenarios 7 / 30

Thrombus imaging with AI for patients stratification Materials Datasets CHSF
MATAR MATAR2 FOCH Num. of annotated pat. 65 125 156 43 Multicentric ✓ Multiequipement ✓ Recanalization ✓ Only proximal ✓ Only distal ✓ ✓ Lesion volume (mL) 31.77 5.01 5.97 - Thrombi volume (mL) 0.23 0.08 0.09 0.31 Proximal vs distal: Factor of 6 8 / 30

Thrombus imaging with AI for patients stratification Materials Datasets: MRI
modalities for stroke detection Fig 2: B0 Fig 3: ADC Fig 4: DWI Fig 5: SWAN Fig 6: PHASE 9 / 30

Thrombus imaging with AI for patients stratification Methods Methodology: Proposed
solutions Multicentric Robustness for missing modalities → Gradual modality dropout ROI Characterization → UpAttLLSTM for thrombus segmentation Stroke evolution assesment → Classification methods using MRI radiomics 10 / 30

Thrombus imaging with AI for patients stratification Methods Multicentric robustness:
Modality incompleteness (a) SWAN (b) PHASE (c) DWI (d) SWI (e) No PHASE (f) DWI 11 / 30

Gradual modality dropout Each modality is multiplied by a value rj before entering the segmentation model during the training phase, which regulates the degree of modality dropout. 12 / 30

Modality incompleteness (g) DWI with rj = 1 (h) DWI with rj = 0.75 (i) DWI with rj = 0.5 (j) DWI with rj = 0.25 Fig 7: Gradual modality dropout examples using DWI 13 / 30

Gradual modality dropout rj ∼ Bernouilli(p) rj = 1, if rj = 1 gj (t), otherwise x = r ⊙ x y = F(x) (1) gj (t) =            0.75, if t < 0.25T 0.5, if t < 0.5T 0.25, if t < 0.75T 0, otherwise (2) 14 / 30

Thrombus imaging with AI for patients stratification Methods ROI: Thrombus
segmentation Fig 8: Thrombus is the cause of the lesion 15 / 30

segmentation Fig 9: UpAttLLSTM. 16 / 30

segmentation: Recurrent Logic Long Short-Term Memory Fig 10: LLSTM. 17 / 30

Thrombus imaging with AI for patients stratification Methods Stroke evolution
assesment: Radiomics1 based on ROI 1Van Griethuysen et al., “Computational radiomics system to decode the radiographic phenotype”. 18 / 30

Thrombus imaging with AI for patients stratification Methods Metrics Accuracy
= TP + TN TP + TN + FP + FN (↑) (3) AUC (↑), Precision (↑). Recall (↑) 19 / 30

Thrombus imaging with AI for patients stratification Methods Metrics Accuracy
= TP + TN TP + TN + FP + FN (↑) (3) AUC (↑), Precision (↑). Recall (↑) Dice(ˆ y, y) = 2ˆ yn ∩ yn ˆ yn ∪ yn (↑) (4) FP (↓), FN (↓) and Detection rate (↑) 19 / 30

Thrombus imaging with AI for patients stratification Results ROI: Thrombus
segmentation using UpAttLLSTM Input modalities: DWI, SWAN, PHASE, SWI Datasets: CHSF, MATAR/MATAR2, FOCH 1 SOTA comparison: nnUnet, CLSTM, LLSTM, Attention, Upasampling, Post-processing 2 Multicentric evaluation: SWI ↔ SWAN under unseen center (FOCH) using gradual modality dropout 20 / 30

segmentation: SOTA comparison nnUnet, CLSTM, LLSTM, AttCLSTM, AttLLSTM, Published in MIDL20242 Monocentric evaluation (CHSF (P) and MATAR (D)) Model Post False Positives (↓) False Negatives (↓) Dice (↑) Det. (↑) # Param (↓) Count Size Count Size (millions) P | D P | D P | D P | D P | D P | D nnUnet 0.6 | 1.1 37.1 | 222.4 0.4 | 0.5 373.6 | 125.3 0.46 | 0.45 0.7 | 0.7 ≃ 30 CLSTM 3.8 | 2.6 178.1 | 153.5 0.2 | 0.4 0.2 | 28.1 0.39 | 0.33 1.0 | 0.81 ≃ 2 LLSTM 0.8 | 1.7 111.9 | 34.9 0.2 | 0.6 0.2 | 136.4 0.48 | 0.36 1.0 | 0.64 ≃ 1 AttCLSTM 1.2 | 1.6 43.81 | 99.3 0.3 | 0.4 141.1 | 100.7 0.41 | 0.38 0.9 | 0.72 ≃ 3 AttLLSTM 1.0 | 0.6 49.8 | 63.3 0.2 | 0.3 0.2 | 79.4 0.55 | 0.54 1.0 | 0.91 ≃ 2 2Vargas-Ibarra, Vigneron, Garcia-Salicetti, et al., “A recurrent network for segmenting the thrombus on brain MRI in patients with hyper-acute ischemic stroke”. 21 / 30

segmentation: SOTA comparison nnUnet, CLSTM, LLSTM, AttCLSTM, AttLLSTM, Published in MIDL20242 Monocentric evaluation (CHSF (P) and MATAR (D)) Model Post False Positives (↓) False Negatives (↓) Dice (↑) Det. (↑) # Param (↓) Count Size Count Size (millions) P | D P | D P | D P | D P | D P | D nnUnet 0.6 | 1.1 37.1 | 222.4 0.4 | 0.5 373.6 | 125.3 0.46 | 0.45 0.7 | 0.7 ≃ 30 CLSTM 3.8 | 2.6 178.1 | 153.5 0.2 | 0.4 0.2 | 28.1 0.39 | 0.33 1.0 | 0.81 ≃ 2 LLSTM 0.8 | 1.7 111.9 | 34.9 0.2 | 0.6 0.2 | 136.4 0.48 | 0.36 1.0 | 0.64 ≃ 1 AttCLSTM 1.2 | 1.6 43.81 | 99.3 0.3 | 0.4 141.1 | 100.7 0.41 | 0.38 0.9 | 0.72 ≃ 3 AttLLSTM 1.0 | 0.6 49.8 | 63.3 0.2 | 0.3 0.2 | 79.4 0.55 | 0.54 1.0 | 0.91 ≃ 2 UpAttLLSTM 2.1 | 2.2 34.2 | 56.9 0.1 | 0.2 0.1 | 62.4 0.62 | 0.58 1.00 | 0.93 ≃ 2 UpAttLLSTM ✓ 0.3 | 0.1 7.8 | 12.5 0.1 | 0.3 0.1 | 65.4 0.65 | 0.65 1.00 | 0.86 ≃ 2 2Vargas-Ibarra, Vigneron, Garcia-Salicetti, et al., “A recurrent network for segmenting the thrombus on brain MRI in patients with hyper-acute ischemic stroke”. 21 / 30

segmentation: SOTA comparison nnUnet, CLSTM, LLSTM, AttCLSTM, AttLLSTM, Published in MIDL20242 Monocentric evaluation (CHSF (P) and MATAR (D)) Model Post False Positives (↓) False Negatives (↓) Dice (↑) Det. (↑) # Param (↓) Count Size Count Size (millions) P | D P | D P | D P | D P | D P | D nnUnet 0.6 | 1.1 37.1 | 222.4 0.4 | 0.5 373.6 | 125.3 0.46 | 0.45 0.7 | 0.7 ≃ 30 CLSTM 3.8 | 2.6 178.1 | 153.5 0.2 | 0.4 0.2 | 28.1 0.39 | 0.33 1.0 | 0.81 ≃ 2 LLSTM 0.8 | 1.7 111.9 | 34.9 0.2 | 0.6 0.2 | 136.4 0.48 | 0.36 1.0 | 0.64 ≃ 1 AttCLSTM 1.2 | 1.6 43.81 | 99.3 0.3 | 0.4 141.1 | 100.7 0.41 | 0.38 0.9 | 0.72 ≃ 3 AttLLSTM 1.0 | 0.6 49.8 | 63.3 0.2 | 0.3 0.2 | 79.4 0.55 | 0.54 1.0 | 0.91 ≃ 2 UpAttLLSTM 2.1 | 2.2 34.2 | 56.9 0.1 | 0.2 0.1 | 62.4 0.62 | 0.58 1.00 | 0.93 ≃ 2 UpAttLLSTM ✓ 0.3 | 0.1 7.8 | 12.5 0.1 | 0.3 0.1 | 65.4 0.65 | 0.65 1.00 | 0.86 ≃ 2 UpAttLLSTM outperforms nnUnet and CLSTM, each of the modules improve the performances 2Vargas-Ibarra, Vigneron, Garcia-Salicetti, et al., “A recurrent network for segmenting the thrombus on brain MRI in patients with hyper-acute ischemic stroke”. 21 / 30

segmentation: Multicentric robustness means that the modality is used in the test and means that it is replaced by a black image. Grad. Test dataset PHASE False Positives (↓) False Negatives (↓) Dice (↑) Det. (↑) mod. Count Size Count Size drop. P | D P | D P | D P | D P | D P | D FOCH (SWAN) 0.8| 23.6 | 1.0 | 334.5 | 0.14 | 0.45 | FOCH (SWI) 1.1| 33.0 | 1.1 | 1007.1 | 0.04 | 0.32 | 22 / 30

segmentation: Multicentric robustness means that the modality is used in the test and means that it is replaced by a black image. Grad. Test dataset PHASE False Positives (↓) False Negatives (↓) Dice(↑) Det. (↑) mod. Count Size Count Size drop. P | D P | D P | D P | D P | D P | D FOCH (SWAN) 0.8 | 23.6 | 1.0 | 334.5 | 0.14 | 0.45 | FOCH (SWI) 1.1 | 33.0 | 1.1 | 1007.1 | 0.04 | 0.32 | ✓ FOCH (SWAN) 4.2 | 101.6 | 0.55 | 81.5 | 0.37 | 0.90 | ✓ FOCH (SWI) 3.9 | 71.5 | 0.55 | 284.1 | 0.28 | 0.80 | Gradual modality dropout improves the performances when a new modality is present 22 / 30

segmentation: Multicentric robustness Segmentation examples Groundtruth UpAttLLSTM UpAttLLSTM +mod drop 23 / 30

segmentation: Multicentric robustness Segmentation examples Groundtruth UpAttLLSTM UpAttLLSTM +mod drop Fig 11: SOTA prediction comparison 23 / 30

segmentation -Conclusions UpAttLLSTM outperforms SOTA proposals Dice of 0.65 detecting more than 90% of thrombi Dice of 0.33 detecting 85% of thrombi in SWI↔ SWAN 24 / 30

Thrombus imaging with AI for patients stratification Results Stroke evolution
assesment Classification models: Random Forest, MLP and Logistic regression Input modalities: DWI, B0, ADC SWAN, PHASE, TOF and FLAIR Datasets: MATAR2 1 Recanalization prediction 25 / 30

assesment: Recanalisation prediction Model Feature selection Groundtruth Prediction Accuracy (↑) AUC (↑) Accuracy (↑) AUC (↑) Logistic Regression P-value 0.760 0.734 0.718 0.754 Fisher Score 0.750 0.846 0.743 0.846 Common 0.723 0.782 0.618 0.727 Random Forest P-value 0.704 0.747 0.627 0.681 Fisher Score 0.803 0.787 0.738 0.799 Common 0.800 0.830 0.759 0.809 MLP P-value 0.701 0.751 0.547 0.753 Fisher Score 0.736 0.796 0.722 0.840 Common 0.720 0.737 0.738 0.800 Table 1: Recanalisation prediction using data imputation 26 / 30

assesment: Conclusions Thrombus shape and MRI modalities information are key Robustness when groundtruth is replaced by predicted mask 76% of accuracy 27 / 30

Thrombus imaging with AI for patients stratification Conclusions and future
works Conclusions Full automatic pipeline that predicts the probability of recanalization after 1h with 76% of accuracy 90% of thrombi detected with 0.65 of Dice 0.32 of Dice in multicentric scenarios with missing modalities detecting 85% of cases thanks to gradual modality dropout 28 / 30

works Future works 1 Under a missing scenario, which model is the best? 2 More modalities 3 Thrombectomy and thrombolysis 29 / 30

works List of publications Sofia Vargas-Ibarra, Vincent Martin Vigneron, Sonia Garcia-Salicetti, et al. “A recurrent network for segmenting the thrombus on brain MRI in patients with hyper-acute ischemic stroke”. In: Proceedings of The 7nd International Conference on Medical Imaging with Deep Learning. Vol. 250. Proceedings of Machine Learning Research. PMLR, Mar. 2024, pp. 657–671 Sofia Vargas-Ibarra, Vincent Martin Vigneron, Hichem Maaref, et al. “Gradual modality dropout for segmenting ischemic stroke lesions in an unseen center with missing modalities”. In: Medical Imaging with Deep Learning. 2025 Sofia Vargas Ibarra et al. “Abstract No : 569 : Automatic Proximal and Distal thrombi segmentation”. In: European Stroke Journal 9 (2024), pp. 3–647 Sofia Vargas Ibarra et al. “Abstract No : 201 : Neurological signs identified by AI related to stroke patient recanalization”. In: European Stroke Journal (2025) Nicolas Chausson et al. “Abstract No : 050 : Early successful recanalization after intravenous thrombolysis with tenecteplase versus alteplase in distal vessel occlusion strokes”. In: European Stroke Journal (2025) Under process: ISBI 2026 Conference, Medical Image Analysis Journal, American Journal of Neuroradiology 30 / 30

Thrombus imaging with AI for patients stratification Nyul normalisation 1
/ 5

Thrombus imaging with AI for patients stratification Nyul normalisation Having
the minimum (s1) and maximum (s2) value for the standard histogram x′ = s1 + x − p1,j p2,j − p1,j (s2 − s1) (5) With all the training images, the rounded mean (µs) is obtained inside the landmarks. With them, the images are normalised as follows: x′ = µs + (x − µi ) s1−µs p1,i −µi , if m1,i ≤ x ≤ µi µs + (x − µi ) s2−µs p2,i −µi , if µi ≤ x ≤ m2,i where µi is the mean inside the landmarks (p1,i , p2,i ) of the image i, where m1,i and m2,i are the extensions of the lower and upper ends of the standard scale. 2 / 5

Thrombus imaging with AI for patients stratification LLSTM Fig 12:
LLSTM. The convolution operation is replaced by the Logic one. This operation reduces the trainable parameters as it is a concatenation of two parts (a1, a2) where smaller convolutions are applied and a pooling layer is included. C2, C4 are 1× 1 convolutional layers 3 / 5

Thrombus imaging with AI for patients stratification ModDrop+ 4 /
5

Thrombus imaging with AI for patients stratification Fisher Score Fisher’s
Score is calculated as the ratio of between-class and within-class variance. A higher Fisher’s Score implies the characteristic is more discriminative and valuable for classification. The score is calculated as follows: Si = (µi,1 − µi,2)2 σ2 i,1 + σ2 i,2 (6) where: µi,1 and µi,2 are the mean values of feature xi for class 1 and 0 respectively, σ2 i,1 and σ2 i,2 are the corresponding variances. 5 / 5

Sofia Vargas Ibarra

Sofia Vargas Ibarra

More Decks by S³ Seminar

Other Decks in Research

Featured

Transcript