Reading chest radiographs for pneumoconiosis by computer

Area of Science:

• Occupational Medicine
• Radiology
• Computer Science

Background:

• Pneumoconiosis diagnosis from radiographs is crucial for epidemiological studies.
• Objective and consistent measurement of pneumoconiosis severity is challenging with human observers.
• Existing automated methods have limitations in accuracy and classification.

Purpose of the Study:

• To describe and assess a novel computer program for measuring simple pneumoconiosis in radiographs.
• To evaluate the program's performance against human expert readings and existing automated techniques.
• To explore the potential of the program for automated epidemiological work.

Main Methods:

• A computer program was developed to quantify density unevenness in 3-6 mm grids of lung radiographs.
• A 'computed score' was derived from the mean diversity across relevant lung fields.
• The computed score was correlated with 'radiological scores' from 11 human observers on 36 films.

Main Results:

• The computer program achieved a correlation of 0.88 with the average radiological score.
• The program successfully classified films into International Labour Office (ILO) U/C categories with 80% accuracy.
• This classification accuracy surpasses reported rates of 45%-65% for other automated methods.

Conclusions:

• The developed computer program shows significant promise for the objective measurement of pneumoconiosis.
• The program's high classification accuracy suggests its potential as a basis for automated epidemiological studies.
• Further validation with larger, diverse film sets is recommended to establish its utility in routine practice.