Deep learning for detecting midline shift on brain CT: a retrospective validation study using multinational data and quantitative measurements

Congress:

ECR 2022

Poster Number:

C-11958

Type:

Scientific Exhibit

Keywords:

Artificial Intelligence, Head and neck, Neuroradiology brain, CAD, CT, Neural networks, Computer Applications-Detection, diagnosis, Acute, Haemorrhage

Authors:

J. Xu, R. Du, C. Mak, J. Abrigo, E. Tsougenis, J. Chan, Q. Dou, P. A. Heng

DOI:

10.26044/ecr2022/C-11958

DOI-Link:

https://dx.doi.org/10.26044/ecr2022/C-11958

Purpose

The degree of midline shift (MLS) is one of the most vital indicators of neurosurgical prognosis that can be associated with other signs of increased intracranial pressure such as brain edema, herniation, and hydrocephalus [1, 2]. Non-contrast CT is the standard imaging modality for evaluating midline shift [3]. Previous research [4] on MLS detection in brain CT has been limited to cross-national validation and quantification. We aim to develop deep learning-based algorithms to detect midline shifts with quantitative measurements and validate the algorithm using multinational...

Methods and materials

Study design and imaging datasets We retrospectively collected 1,500 brain CT scans (denoted as HK1500 dataset) from two hospitals through Hong Kong Hospital Authority (HA) AI Lab. The cohort included adult patients between 2014 and Dec 2018, with relevant ethical approvals obtained. The imaging data were selected based on their associated reports (Fig. 2). First, 269,235 clinical radiology reports of brain CT cases were collected and classified using a rule-based natural language processing (NLP) algorithm, which was used to identify the presence of MLS. Then,...

Results

[Fig 7] The HK1500 dataset consisted of 1,489 brain CT scans from two centers, i.e., 1,117 from Prince of Wales Hospital (PWH) and 372 from Queen Elizabeth Hospital (QEH). 1,500 brain CT scans were included in the initial data retrieval process. Of these, 11 scans were excluded based on our exclusion criteria due to the non-availability of non-contrast 5-mm axial CT series (n=7), and the scan did not cover the whole brain (n=4). The cohort characteristics are summarized in Table 1. 1,090 (73%) out of...

Conclusion

In this study, we show that deep learning algorithms can accurately detect midline shifts in brain CT of patients from multiple centers and regions, with a high AUC over 0.9 but a slight performance drop across cohorts. We were also able to quantify the degrees of MLS when the measurements are available and achieve high sensitivity and specificity, which can potentiallyhelp patient monitoring. Our multinational validation showcases the generalization ability and issues that can facilitate the development of deep learning-based methods on clinical applications. We...

Personal information and conflict of interest

J. Xu: Nothing to disclose R. Du: Nothing to disclose C. Mak: Nothing to disclose J. Abrigo: Nothing to disclose E. Tsougenis: Nothing to disclose J. Chan: Nothing to disclose Q. Dou: Nothing to disclose P. A. Heng: Nothing to disclose

References

1. Hemphill III, J. Claude, et al. "Guidelines for the management of spontaneous intracerebral hemorrhage: a guideline for healthcare professionals from the American Heart Association/American Stroke Association." Stroke 46.7 (2015): 2032-2060. 2. Bullock, M. Ross, et al. "Surgical management of traumatic parenchymal lesions." Neurosurgery 58.suppl_3 (2006): S2-25. 3. Gruen, Peter. "Surgical management of head trauma." Neuroimaging Clinics 12.2 (2002): 339-343. 4. Chilamkurthy, Sasank, et al. "Deep learning algorithms for detection of critical findings in head CT scans: a retrospective study." The Lancet 392.10162 (2018): 2388-2396....

Fig 1: Study overview. We developed deep learning-based algorithms to detect midline...

Fig 2: Data selection flow. NLP=natural language processing. MLS=midline shift.

Fig 3: Preprocessing overview. a. The preprocessing pipeline. We extracted volume data...

Fig 4: Deep neural network (DNN) architecture. The developed DNN comprises 2D CNN to...

Fig 5: Receiver operating characteristic (ROC) curves for midline shift detection on...

Fig 6: Examples of correct predictions and failure cases by the deep neural network....

Fig 7: Table 1. The characteristics of the collected data of two centers from Hong...

Fig 8: Table 2. Midline shift detection performance for models trained using CQ500 and...

Fig 9: Table 3. Multi-degree midline shift classification performance using HK1500.

Fig 10: Table 4. Midline shift detection performance for models trained using HK1500...