You can try looking at the zero crossings of the second derivative of the S-T segment to find the border between the S and T waves.

Alternatively you might be able to use shape matching by using a small annotated set of reference T-waves and aligning them with the T-wave in the S-T segment by finding the position with maximum cross correlation.

thank you Mr.Michael.

Zero crossings of second derivative suggested by Michael Rooijakkers are too noisy and will produce jitter. His second suggestion of 'shape matching' is better.

First and second derivatives together can tell you shape and curve trends. But, you must take into account for the fact that derivatives intrinsically act as high pass filters and  therefore they reduce the Signal-to-noise ratio. May be it would be necessary to smooth the signal before taking the derivative

This is a difficult problem in general unless you have a specific parametric form for the curve you are looking for. Filligoi's approach seems reasonable. You must definitely smooth the signal to reduce the noise first. You could use a Kalman smoother to do this. The start and end points of the curve need to be identified and the general shape of the curve has to be validated. You will need to develop your own rules of thumb for this based on your data.

I recently developed a set of rules for detecting a curve in 1-D passive sonar data, corresponding to an accelerating or decelerating acoustic source or a constant velocity source passing close to the receiver. This data is very noisy and also has false alarms. A brief summary is contained in the link below.

Conference Paper A High Performance 1-D Hidden Markov Model Tracker for Passi...