fetal-planes-classification-custom-cnn-V1

This model was trained from scratch on ERO26/fetal-planes-classification-zenodo. It achieves the following results on the evaluation set:

• Loss: 0.3586
• Accuracy: 0.9080
• Precision Macro: 0.8620
• Recall Macro: 0.9027
• F1 Macro: 0.8790

Model description

• Developed by: ERO26
• Model Type: Custom Convolutional Neural Network (CNN) for Image Classification
• Task: Multi-class classification of standard fetal ultrasound planes
• Input: Grayscale ultrasound images resized to 128x128 pixels ((1, 128, 128) tensor format)
• Output: Probability logits across the target fetal plane classes
• Architecture Highlights:
  • Feature Extractor: 3 Conv2d blocks (32, 64, and 128 channels) with BatchNorm2d, ReLU, MaxPool2d, and Dropout2d.
  • Classifier Head: Flatten layer, linear projection to 256 hidden units with ReLU and 50% dropout, followed by a final linear layer matching num_classes.
• Loss Function: Cross-Entropy Loss with integrated class weights to handle dataset imbalance.

Intended uses & limitations

Intended Uses

• Clinical Quality Assurance: Assisting sonographers and medical researchers in automatically tagging or verifying standard fetal imaging planes during routine or archive scans.
• Educational Tool: Supporting medical students and trainees in learning standard fetal biometric and anatomical plane identification.

Limitations & Out-of-Scope Use

• Not a Diagnostic Device: This model is intended for research, educational, and assistive prototyping purposes only. It must not be deployed as a standalone diagnostic tool for clinical decision-making.
• Domain Shift: Performance may degrade when evaluating ultrasound images captured using scanner hardware, acoustic settings, or transducers significantly different from the training distribution.

Training and evaluation data

• Dataset: Trained on ERO26/fetal-planes-classification-zenodo.
• Data Splits: The dataset is partitioned into training, validation, and test subsets to ensure robust generalization and unbiased final evaluation.
• Preprocessing: Images are normalized, converted to single-channel grayscale, and uniformly resized to 128x128 pixels before being ingested into the network pipeline.

Training procedure

The model was trained from scratch using the Hugging Face Trainer framework with native Automatic Mixed Precision (AMP) enabled for accelerated convergence. Class frequency imbalances were addressed by incorporating dynamically calculated class weights into the loss function.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.001
• train_batch_size: 4
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9, 0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 50
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Precision Macro	Recall Macro	F1 Macro
0.9073	1.0	317	0.9008	0.6512	0.6269	0.6025	0.5566
0.9162	2.0	634	0.7200	0.7058	0.6486	0.7089	0.6609
0.8177	3.0	951	0.6554	0.8049	0.7461	0.7387	0.7393
0.7520	4.0	1268	0.5613	0.7543	0.6862	0.7465	0.6999
0.6695	5.0	1585	0.5708	0.8038	0.7450	0.7892	0.7591
1.0570	6.0	1902	0.5302	0.7907	0.7236	0.7826	0.7435
0.8156	7.0	2219	0.5057	0.7968	0.7551	0.8177	0.7715
0.6888	8.0	2536	0.4715	0.8160	0.7591	0.8324	0.7835
0.6194	9.0	2853	0.4609	0.8220	0.7636	0.8441	0.7900
0.6261	10.0	3170	0.4942	0.7978	0.7498	0.8299	0.7630
0.5964	11.0	3487	0.4395	0.8413	0.7870	0.8435	0.8039
0.7691	12.0	3804	0.4477	0.8322	0.7760	0.8446	0.8005
0.5971	13.0	4121	0.4079	0.8413	0.7851	0.8580	0.8096
0.6364	14.0	4438	0.4061	0.8413	0.7832	0.8596	0.8076
0.6412	15.0	4755	0.4355	0.8544	0.8021	0.8623	0.8229
0.5242	16.0	5072	0.3783	0.8686	0.8178	0.8573	0.8334
0.6033	17.0	5389	0.3738	0.8625	0.8045	0.8701	0.8287
0.7080	18.0	5706	0.3965	0.8483	0.8013	0.8617	0.8166
0.5689	19.0	6023	0.4015	0.8524	0.8005	0.8724	0.8248
0.4923	20.0	6340	0.3855	0.8574	0.8025	0.8750	0.8244
0.4272	21.0	6657	0.3553	0.8766	0.8264	0.8738	0.8460
0.5221	22.0	6974	0.3597	0.8615	0.8095	0.8804	0.8298
0.4752	23.0	7291	0.3648	0.8857	0.8392	0.8759	0.8528
0.3906	24.0	7608	0.3677	0.8706	0.8189	0.8701	0.8387
0.4001	25.0	7925	0.3202	0.8706	0.8160	0.8867	0.8406
0.4809	26.0	8242	0.3232	0.8797	0.8326	0.8942	0.8549
0.4715	27.0	8559	0.3460	0.8787	0.8299	0.8916	0.8508
0.3758	28.0	8876	0.3383	0.8888	0.8378	0.8879	0.8574
0.3891	29.0	9193	0.3455	0.8857	0.8335	0.8966	0.8577
0.2582	30.0	9510	0.3347	0.8969	0.8455	0.8992	0.8673
0.3163	31.0	9827	0.3198	0.8938	0.8428	0.8891	0.8615
0.3230	32.0	10144	0.3234	0.8868	0.8305	0.8821	0.8511
0.3189	33.0	10461	0.3098	0.9019	0.8545	0.9001	0.8727
0.3142	34.0	10778	0.3434	0.8979	0.8472	0.8947	0.8666
0.3274	35.0	11095	0.3387	0.9070	0.8617	0.8952	0.8759
0.4326	36.0	11412	0.3423	0.9029	0.8576	0.8963	0.8732
0.4480	37.0	11729	0.3586	0.9080	0.8620	0.9027	0.8790
0.2945	38.0	12046	0.3761	0.8969	0.8444	0.8820	0.8608
0.2479	39.0	12363	0.3716	0.8827	0.8321	0.8721	0.8481
0.2763	40.0	12680	0.3532	0.8989	0.8475	0.8950	0.8669
0.3422	41.0	12997	0.3251	0.9039	0.8562	0.8936	0.8715
0.3239	42.0	13314	0.3264	0.8999	0.8475	0.8954	0.8671

Framework versions

• Transformers 5.16.1
• Pytorch 2.11.0+cu128
• Datasets 5.0.1
• Tokenizers 0.23.1