TL;DR
This paper introduces the first method for estimating the post-mortem interval (PMI) directly from forensic iris images. By training deep learning models on NIR, RGB, and multispectral iris data, the researchers demonstrate that iris patterns can provide valuable clues about the time elapsed since death, even in challenging cross-dataset scenarios.
What Problem Does This Paper Address?
Estimating the post-mortem interval (PMI) is a critical but challenging task in forensic science. While traditional methods rely on physical changes in the body (like cooling or decomposition), these can be influenced by environmental factors and are not always precise. The authors propose a novel approach: using iris images, a biometric feature typically used for live identification, to estimate PMI. This is particularly interesting because iris decomposition is correlated with PMI, and an automated method could significantly aid forensic investigations, especially in mass fatality incidents or cold cases. Previous assumptions suggested iris recognition was impossible after death due to rapid degradation, but recent studies hint at a more moderate decay process.
Methods
The core of this work involves training deep learning models to predict PMI from iris images. The researchers experimented with several established CNN architectures (VGG19, DenseNet121, ResNet152, Inception_v3) and a Vision Transformer (ViT). They explored three main approaches:
- Spectral Modalities: Models were trained using only Near-Infrared (NIR) data, only Visible (RGB) data, or a combination of both (multispectral).
- Model Architectures: Standard CNNs were adapted for regression tasks, and for multispectral data, a fusion model combining NIR and RGB embeddings was used.
- Data Balancing for Cross-Dataset Scenarios (S3): To improve generalization to unseen data (from different labs), two techniques were employed:
- S3-real: Up-sampling real data with replacement to balance PMI ranges.
- S3-synthetic: Augmenting the training data with synthetically generated forensic iris images (using StyleGAN2-ADA) to create a uniform PMI distribution.
Rigorous cross-validation was performed across three scenarios: sample-disjoint (S1), subject-disjoint (S2), and dataset-disjoint (S3), with S3 being the most realistic.
Datasets
The study utilized a combined dataset of 8,064 NIR and 7,215 RGB forensic iris images from 348 deceased subjects. This dataset was compiled from publicly available sources:
- Warsaw BioBase Post Mortem Iris v2.0: 1,787 RGB and 1,200 NIR images from 37 individuals, collected in a mortuary setting with controlled temperature.
- Warsaw BioBase Post Mortem Iris v3.0: 785 RGB and 1,094 NIR images from 42 individuals, with similar environmental conditions to v2.0.
- NIJ-2018-DU-BX-0215: 4,643 RGB and 5,770 NIR images from 269 individuals, collected during routine operations of a medical examiner's office.
The Warsaw datasets were combined for analysis, and the NIJ dataset was used separately.
Key Insights and Ideas
- Iris as a PMI Indicator: The study definitively shows that iris patterns, even after death, contain features indicative of the post-mortem interval. This challenges previous assumptions about the rapid degradation of iris texture.
- Multispectral Advantage: Combining NIR and RGB data (multispectral) consistently yielded better PMI estimation results, especially in more challenging scenarios. This suggests that different spectral bands capture complementary decomposition-related information.
- Synthetic Data for Generalization: Augmenting training data with synthetically generated forensic iris images (S3-synthetic) significantly improved model generalization, particularly in the cross-dataset scenario (S3). This is crucial for building robust models applicable to real-world forensic situations where data may come from diverse sources.
- Dataset Diversity Matters: The study highlights that training on datasets with a wider range of PMIs (like the NIJ dataset) leads to better generalization than training on datasets with narrower, more uniform PMI distributions (like the Warsaw dataset). This is evident in the cross-dataset evaluation where models trained on NIJ performed better on Warsaw than vice-versa.
Key Results
- Scenario S1 (Sample-disjoint): The best performance was achieved with multispectral data, yielding a Mean Absolute Error (MAE) of approximately 3.5 hours.
- Scenario S2 (Subject-disjoint): Performance degraded, with the best multispectral model achieving an MAE of around 17.5 hours.
- Scenario S3 (Dataset-disjoint): This most realistic scenario showed the highest errors. The best multispectral model with S3-synthetic balancing achieved an MAE of approximately 45.77 hours.
- NIR vs. RGB: While both modalities provided useful information, multispectral data generally outperformed single-band approaches. DenseNet121 and ResNet152 were often top performers for NIR and multispectral data, respectively, in the less challenging scenarios.
- Synthetic Data Impact: In the S3 scenario, synthetic data balancing reduced the MAE by roughly 3 days (77.66 hours) down to approximately 2 days (45.77 hours) for multispectral data.
Limitations and Future Work
- Contextual Factors: The study acknowledges that relying solely on iris images overlooks crucial contextual factors like ambient temperature, humidity, and other post-mortem changes in tissues, which could improve PMI estimation.
- Biological Bias: Other biological factors (gender, age, death reason, diet) might introduce bias into the models.
- Dataset Limitations: Despite rigorous cross-validation, the dataset size and diversity might still limit generalization.
- Future Work: The authors suggest incorporating additional contextual information and leveraging advancements in generative models for more accurate synthetic data generation. They also propose using the estimated PMI to condition post-mortem iris recognition methods, adapting classifiers based on the degree of iris decomposition.
Summary updated