A deep-learning algorithm effectively detects pathologic lymph node extranodal extension (ENE) on preoperative CT scans in patients with laryngeal and hypopharyngeal squamous cell cancer (LHSCC), researchers have reported.
In fact, the algorithm bested head and neck cancer specialists for this task in the area under the receiver operating curve (AUC) and sensitivity measures, wrote a team led by Na Shen, MD, of Fudan University in Xiamen, China. The findings were published on January 13 in Radiology.
“We developed a new deep-learning diagnostic tool, DeepENE, that accurately detected extranodal extension on preoperative CT scans in patients with laryngeal and hypopharyngeal squamous cell cancer and outperformed head and neck cancer specialists,” the group noted.
Although accurate presurgical identification of ENE on CT imaging is essential for determining effective treatment for LHSCC, human interpretation of the condition hasn’t proven reliable or reproducible, the investigators explained. Shen and colleagues sought to assess the diagnostic performance of a deep-learning tool they created called DeepENE for identifying metastatic ENE and ENE lymph nodes on preoperative CT scans in patients with LHSCC.
The team conducted a study that included data from 289 patients with LHSCC with 1,954 pathologically confirmed lymph nodes between April 2011 and August 2022. The group used this data to establish training, validation, and internal test sets for DeepENE; the reference standard was lymph nodes segmented on CT scans and labeled for metastasis as well as ENE status based on pathologic findings.
The investigators tested DeepENE using three external test sets made up of patients with LHSCC from three different institutions and one external test set made up of patients with oral squamous cell carcinoma. They evaluated the performance of the algorithm using the AUC measure and compared its results to those of five head and neck cancer specialists.
DeepENE showed an AUC of 0.93 for ENE diagnosis in the internal test set, the researchers reported. They also found the following:
|
Deep-learning algorithm for identifying ENE compared with radiologist readers |
||
|
Measure |
Radiologist readers |
DeepENE |
|
AUC |
||
| External test set 1 |
0.85 |
0.96 |
| External test set 2 |
0.66 |
0.87 |
| External test set 3 |
0.71 |
0.9 |
|
Sensitivity |
||
| External test set 1 |
77% |
97% |
| External test set 2 |
36% |
78% |
| External test set 3 |
46% |
87% |
|
Specificity |
||
| External test set 1 |
93% |
90% |
| External test set 2 |
96% |
90% |
| External test set 3 |
96% |
87% |
|
Accuracy |
||
| External test set 1 |
92% |
91% |
| External test set 2 |
89% |
89% |
| External test set 3 |
93% |
87% |
Going forward, the authors plan to continue refining DeepENE.
“Our deep-learning model will be prospectively evaluated as a second-opinion assist tool for precise and personalized pretreatment assessment in patients with LHSCC,” they concluded.
The complete study can be accessed here.