Diagnosing pulmonary embolism by integrating patient-level diagnosis and embolus-level detection
View Patent ↗Improved pulmonary embolism (PE) detection may be obtained through computer aided-diagnosis. In particular, PE detection may be accomplished through patient-level diagnosis, embolus-level detection, or a combination of the two. Patient-level diagnosis operates to quickly exclude non-PE patients and dispatch PE-patients to treatment. Embolus-level detection operates to localize individual emboli to support personalized medicine via risk stratification. Multiple instance-based learning (MIBL) classification at the patient level explores the key observation that once any TP candidate of a patient is classified as positive, the patient is identified as PE positive. That is, MIBL focuses on correct classification of patients rather than individual candidates, to effectively and rapidly distinguish between PE patients and non-PE patients.
1. A method, comprising:
receiving candidates for classification as true positives and false positives, wherein the candidates are associated with one of a plurality of groups;
determining a classification boundary for the true positives and the false positives;
calculating hinge errors for the received candidates based, at least in part, on the determined classification boundary;
setting a portion of the calculated hinge errors to zero, wherein the portion comprises hinge errors corresponding to candidates on a false positive side of the determined classification boundary belonging to a group of the plurality of groups also having candidates on a true positive side of the determined classification boundary;
determining a shifted classification boundary based, at least in part, on the hinge errors after setting the portion of the calculated hinge errors to zero; and
identifying as true positives the candidates on a true positive side of the shifted classification boundary.
2. The method of claim 1 , wherein the step of determining a shifted classification boundary comprises applying a support vector machine (SVM) to construct a linear classification function by minimizing the hinge errors.
3. The method of claim 2 , wherein the step of constructing the linear classification function comprises constructing the linear classification function such that a minimum hinge error for candidates associated with each group of the plurality of groups is zero.
4. The method of claim 2 , wherein the step of minimizing the hinge errors comprises solving the optimization problem defined by
min w,ξ,λ γ∥w∥ 1 +Σ p=1 n (Σ i∈I p + λ i ξ i )+Σ i∈I −ξ i
s.t. w T x i +b≥ 1−ξ i , i∈I + ,
w T x i +b≤− 1+ξ i , i∈I − ,
ξ i ≥0 , i= 1 , . . . , l,
Σ i∈I p + λ i =1, λ i ≥0, i∈I p + , p= 1, . . . , n.
5. The method of claim 2 , wherein the linear classification function is of the form w T x+b and wherein the hinge error is defined as max{0,1−y (w T x−b)}.
6. The method of claim 1 , further comprising aggregating multiple classifiers to reduce a variance of the determination of the classification boundary over multiple received candidates.
7. The method of claim 1 , wherein the plurality of groups corresponds to a plurality of patients.
8. The method of claim 7 , further comprising identifying a portion of the plurality of patients having a pulmonary embolism, wherein each patient of the identified portion has at least one candidate identified as a true positive.
9. The method of claim 7 , further comprising:
building a second set of candidates by selecting, from the received candidates, candidates corresponding to the portion of the plurality of patients identified as having a pulmonary embolism, wherein the second set of candidates corresponds to a plurality of emboli;
determining a second classification boundary for the second set such that a minimum hinge error for candidates of the second set of candidates associated with each embolism of the plurality of emboli is zero; and
identifying as true positives the candidates on a true positive side of the second classification boundary.
10. The method of claim 9 , wherein the step of determining the second classification boundary comprises determining a second shifted classification boundary comprising: determining a classification boundary for true positives and false positives; calculating hinge errors for the second set of candidates based, at least in part, on the determined classification boundary; setting a portion of the calculated hinge errors to zero, wherein the portion comprises hinge errors corresponding to candidates on a false positive side of the determined classification boundary corresponding to embolisms also having candidates on a true positive side of the determined classification boundary; determining the second shifted classification boundary based, at least in part, on the hinge errors after setting the portion of the calculated hinge errors to zero.
11. The method of claim 1 , wherein the plurality of groups corresponds to a plurality of pulmonary emboli.
12. A computer program product, comprising:
a non-transitory computer readable medium comprising code to perform the steps of:
receiving candidates for classification as true positives and false positives, wherein the candidates are associated with one of a plurality of groups;
determining a classification boundary for the true positives and the false positives;
calculating hinge errors for the received candidates based, at least in part, on the determined classification boundary;
setting a portion of the calculated hinge errors to zero, wherein the portion comprises hinge errors corresponding to candidates on a false positive side of the determined classification boundary belonging to a group of the plurality of groups also having candidates on a true positive side of the determined classification boundary;
determining a shifted classification boundary based, at least in part, on the hinge errors after setting the portion of the calculated hinge errors to zero; and
identifying as true positives the candidates on a true positive side of the shifted classification boundary.
13. The computer program product of claim 12 , wherein the plurality of groups corresponds to a plurality of patients.
14. The computer program product of claim 13 , wherein the medium further comprises code to perform the step of identifying a portion of the plurality of patients having a pulmonary embolism, wherein each patient of the identified portion has at least one candidate identified as a true positive.
15. The computer program product of claim 12 , wherein the plurality of groups corresponds to a plurality of pulmonary emboli.
16. An apparatus, comprising:
a memory; and
a processor coupled to the memory, wherein the processor is configured to perform the steps of:
receiving candidates for classification as true positives and false positives, wherein the candidates are associated with one of a plurality of groups;
determining a classification boundary for the true positives and the false positives;
calculating hinge errors for the received candidates based, at least in part, on the determined classification boundary;
setting a portion of the calculated hinge errors to zero, wherein the portion comprises hinge errors corresponding to candidates on a false positive side of the determined classification boundary belonging to a group of the plurality of groups also having candidates on a true positive side of the determined classification boundary;
determining a shifted classification boundary based, at least in part, on the hinge errors after setting the portion of the calculated hinge errors to zero; and
identifying as true positives the candidates on a true positive side of the shifted classification boundary.
17. The apparatus of claim 16 , wherein the plurality of groups corresponds to a plurality of patients.
18. The apparatus of claim 17 , wherein the processor is further configured to perform the step of identifying a portion of the plurality of patients having a pulmonary embolism, wherein each patient of the identified portion has at least one candidate identified as a true positive.
19. The apparatus of claim 16 , wherein the plurality of groups corresponds to a plurality of pulmonary emboli.