Placenta Accreta Spectrum Detection

• Placenta Accreta Spectrum (PAS) is a group of obstetric disorders characterized by abnormal placental invasion into the uterine wall, increasing maternal morbidity and mortality.
• Detection methods integrate ultrasound and MRI imaging with deep learning algorithms, including CNNs and transformer models, to enhance diagnostic accuracy.
• Clinical studies show that multimodal fusion models achieve up to 90.5% accuracy and an AUC of 0.902, significantly outperforming unimodal approaches in reducing surgical risks.

Placenta Accreta Spectrum (PAS) encompasses a spectrum of obstetric disorders characterized by varying degrees of abnormal trophoblastic invasion of the uterine wall. This condition is a significant contributor to maternal morbidity and mortality due to its association with severe postpartum hemorrhage, emergency hysterectomy, and life-threatening complications. Accurate prenatal detection of PAS and its subtypes—placenta accreta (PA), increta (PI), and percreta (PP)—is essential for guiding clinical decision-making, optimizing peripartum resource allocation, and reducing adverse outcomes. Detection methodologies have evolved rapidly with advances in medical imaging and artificial intelligence, particularly leveraging multimodal imaging data and complex deep learning-based pattern recognition frameworks.

1. Clinical Context and Motivations

PAS describes aberrant adherence of the placenta to the myometrium, ranging from superficial attachment (PA) to deep myometrial invasion (PI), and, in the most severe form, transmural progression with serosal or extrauterine organ involvement (PP). The clinical imperative for reliable, early, and accurate PAS detection arises from the direct correlation between diagnostic confidence and reduced rates of catastrophic hemorrhage, surgical complications, and neonatal compromise. Imaging modalities employed include two-dimensional ultrasound (US) for real-time interface and vascularity assessment, and magnetic resonance imaging (MRI) for high-contrast delineation of placental invasion and regional anatomy (Jiang et al., 23 May 2025, Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).

While US is widely accessible and operator-dependent, MRI provides comprehensive volumetric evaluation, yet is resource-intensive and subject to interpretive variability, motivating the integration of automated, reproducible computational detection tools (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025). Recent research emphasizes multimodal fusion to harness the strengths of both modalities and overcome limitations inherent to each.

2. Imaging Datasets and Preprocessing

PAS detection research utilizes well-curated, large-scale imaging datasets, with standardized acquisition and quality control. Exemplary datasets include 1,133 T2-weighted MRI volumes (853 normal, 280 PAS) and 983 2D US images (676 normal, 307 PAS), as well as 160 patient-matched MRI-US pairs for multimodal analysis (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).

Imaging data preprocessing protocols comprise the following standardized steps:

• MRI pipeline: DICOM to NIfTI conversion, canonical reorientation, cubic interpolation to fixed 3D resolution (e.g., $128 \times 128 \times 64$), zero-padding to preserve structure, and per-volume min–max intensity normalization ($$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$) (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).
• US pipeline: Rescaling to $224 \times 224$, conversion to 3-channel RGB, global min–max normalization, with label smoothing ($\epsilon = 0.1$) in certain settings (Ali et al., 31 Dec 2025).
• Augmentation: On-the-fly random flips, rotations ($\pm10^\circ$ or multiples of 90°), random scaling (MRI: $1.1$–$1.3\times$), and intensity jitter to improve class balance and robustness (Jiang et al., 23 May 2025, Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).

Case-control splits are stratified at the patient level, typically in 70:10:20 (train:validation:test) or 60:15:25 (for paired data), with class balancing handled via oversampling and augmentation (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).

3. Deep Learning Frameworks for PAS Detection

State-of-the-art PAS detection employs deep convolutional and transformer-based neural architectures, with progressive movement toward hybrid and multimodal models.

3.1 Anatomy-Guided Multitask CNNs

A multitask convolutional neural network (CNN) architecture enables end-to-end MRI-based detection and PAS subtype classification (Jiang et al., 23 May 2025). The network consists of:

• A main classification branch (backbone) utilizing a four-block ResNet encoder (channels: $16 \to 256$).
• An anatomy-guided segmentation branch, U-Net style, for soft region-of-interest masking (placenta and serosal layer).

The model accepts a $10$-slice stack ($448 \times 448 \times 10$) of sagittal T2 MRI and yields four-class softmax probabilities (non-PAS, PA, PI, PP), as well as pixelwise anatomical segmentations. Feature fusion is realized via element-wise attention weighting using the segmentation mask ($$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$0) at the classification bottleneck.

A multitask loss combines cross-entropy classification and binary cross-entropy segmentation, $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$1, with optimal weighting achieved at $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$2 (Jiang et al., 23 May 2025).

3.2 Hybrid 3D CNN-Transformer Models

Hybrid models integrating 3D DenseNet121 and 3D Vision Transformer (ViT) modules address binary PAS detection on whole MRI volumes (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025). The MRI volume is processed in parallel by:

• 3D DenseNet121: Local feature extraction via dense connectivity, bottleneck, and transition layers.
• 3D ViT: Global spatial context modeling, dividing the volume into non-overlapping $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$3 patches, projecting each patch to a 768-dimensional token, and feeding the token sequence into a 12-layer transformer encoder.

The concatenated DenseNet and ViT embeddings ($$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$4) are classified by a multilayer perceptron with softmax output. The mathematical backbone for ViT self-attention is the scaled dot-product attention: $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$5.

A standard binary cross-entropy loss is adopted, with learning rate $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$6, Adam optimizer, and ReduceLROnPlateau scheduling (Ali et al., 21 Dec 2025).

3.3 Multimodal MRI–US Fusion Models

PAS detection performance is further enhanced by intermediate feature-level fusion of MRI and US through multimodal deep networks (Ali et al., 31 Dec 2025). The framework uses:

• MRI encoder: 3D DenseNet121–ViT, output embedding $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$7.
• US encoder: 2D ResNet50, output embedding $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$8.
• Fused representation: concatenation $$X_\text{norm}(i,j,l) = \frac{X(i,j,l) - \min X}{\max X - \min X}$$9, classified with an MLP (ReLU, dropout 0.5, sigmoid output).

This configuration demonstrates robust synergy, outperforming unimodal networks on balanced test cohorts.

4. Diagnostic Performance and Comparative Analysis

Quantitative evaluation is conducted with stratified test sets and appropriate metrics (accuracy, AUC, precision, recall, F1), reported as means across multiple random seeds or cross-validation folds.

Summary table: PAS detection approaches and comparative performance

Model	Input	Accuracy (%)	AUC	Task	Reference
Anatomy-Guided Multitask CNN	10-slice MRI stack	48.2 (macro, 4-way)	0.8015 (macro)	4-class	(Jiang et al., 23 May 2025)
DenseNet121–ViT (MRI)	3D MRI volume	84.3 ± 1.3	0.842 ± 0.012	PAS vs. normal	(Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025)
ResNet50 (US)	2D US image	86.9 ± 1.7	0.856 ± 0.022	PAS vs. normal	(Ali et al., 31 Dec 2025)
Multimodal Fusion (MRI+US)	MRI volume + US image	90.5 ± 1.9	0.902 ± 0.024	PAS vs. normal	(Ali et al., 31 Dec 2025)

• The anatomy-guided multitask CNN yields a macro-AUC of 0.8015 and accuracy of 48.2% in a challenging four-class (non-PAS/PA/PI/PP) task (Jiang et al., 23 May 2025).
• The hybrid DenseNet121–ViT achieves 84.3% ± 1.3 accuracy and AUC 0.842 ± 0.012 on binary MRI PAS detection (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).
• The US-only ResNet50 model achieves 86.9% ± 1.7 accuracy and AUC 0.856 ± 0.022 (Ali et al., 31 Dec 2025).
• The multimodal fusion model achieves 90.5% ± 1.9 accuracy and AUC 0.902 ± 0.024, significantly outperforming MRI-only and US-only approaches, with statistical significance confirmed by repeated-measures ANOVA and FDR-corrected t-tests ($224 \times 224$0 vs. MRI, $224 \times 224$1 vs. US) (Ali et al., 31 Dec 2025).

Ablation confirms that each imaging modality contributes unique discriminative information, and that simple embedding concatenation is sufficient for effective joint representation (Ali et al., 31 Dec 2025).

5. Training, Optimization, and Evaluation Protocols

Model training across all architectures follows rigorous protocols:

• Optimizer: Adam ($224 \times 224$2), weight decay $224 \times 224$3 in some models.
• Learning-rate scheduling: ReduceLROnPlateau for validation loss.
• Batch size: Typically 8 for 3D models; 16 for 2D stack models.
• Epochs: 100–200, with early stopping after 500 epochs (if no improvement).
• Evaluation: Accuracy, AUC, precision, recall, and F1-score; five-fold cross-validation or five independent runs to capture statistical variability (Jiang et al., 23 May 2025, Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).

Class balancing is achieved via PAS oversampling in training sets, and segmentations for anatomical supervision are annotated by radiologists or generated with semi-automatic nnU-Net plus manual refinement (Jiang et al., 23 May 2025).

6. Clinical and Practical Implications

Automated PAS detection frameworks are positioned to transform clinical workflows:

• Radiologist decision support: Models present real-time risk stratification, probability maps, and anatomical segmentations. For the multitask CNN, inference time is $224 \times 224$40.15 s per patient, with a $224 \times 224$5 minute reduction in manual region-of-interest annotation (Jiang et al., 23 May 2025).
• Diagnostic efficiency and quality: One-stage multiclass models reduce error accumulation and expedite the diagnostic pipeline by $224 \times 224$650% compared to cascaded approaches (Jiang et al., 23 May 2025).
• Multidisciplinary care: Early identification of PAS, especially accurate parsing of PA/PI/PP, informs the need for resources such as blood products and surgical planning (e.g., hysterectomy preparedness) (Jiang et al., 23 May 2025).
• Reduction in inter-observer variability: Machine learning models standardize detection, decreasing interpretive discrepancies (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).
• Limitations and future directions: Current studies are single-center, with a need for multi-institutional validation to ensure generalizability. Expansion to larger, more diverse multimodal datasets is required, as well as exploration of attention-based fusion, explainability (e.g., Grad-CAM), and clinical-data integration (Ali et al., 31 Dec 2025).

7. Directions for Future Research

Identified avenues for further advancement include:

• Multi-center, protocol-agnostic validation to mitigate scanner and demographic biases (Ali et al., 21 Dec 2025, Ali et al., 31 Dec 2025).
• Integration with clinical and biochemical data (risk factors, serum markers) for holistic decision support (Ali et al., 31 Dec 2025).
• Development of interpretable AI with attention visualization, region importance mapping, and clinician-centric model explanations (Ali et al., 21 Dec 2025).
• Exploration of dynamic, attention-weighted fusion for multimodal inputs to improve discriminability and robustness (Ali et al., 31 Dec 2025).
• Segmentation-guided classification frameworks focusing directly on the uterine-placental junction to further improve specificity (Ali et al., 31 Dec 2025).

A plausible implication is that, as multimodal datasets grow and fusion methodologies mature, PAS detection models will become integral to obstetric imaging protocols, enabling reproducible, high-accuracy, real-time diagnosis and cascade prevention of severe PAS-related complications.