I've been working on building instruments with a photodiode and a high gain TIA. I started out with a 100M feedback resistor, which actually worked pretty well. I have a heater serpentine trace on an inner layer under the photodiode and analog stage, with a thermistor placed near the detector. I have a PID temp control loop keeping that part of the PCB at 40C, with a PWM'd FET from my microcontroller putting power into the heater trace. There is a groundplane between the heater trace and the analog parts. The high gain parts and the photodiode traces are all on the top side, no vias. Temp controlling all the parts is pretty important, and even then, I need to use resistors with <100ppm tempco. Everything is in a metal light tight box. I've made almost 100 of these, and at 100M feedback the performance is good. They do have to be calibrated using a precision blackbody source, but after that they are extremely accurate. I am only interested in <1Hz signal though, so not sure how this would work out for higher bandwidths.
For other system level reasons, the light flux has been cut to 1/4 of what it was, so now I'm up to 400M feedback resistors, and having some trouble with measurement stability. So I'm thinking of moving to an integrator to see how that works. Another benefit of the integrator is that you can change the gain by changing the integration time. Right now I use an SMT reed relay to switch in a lower value parallel feedback resistor to change gain, which has some annoyances and is not fast.
If you do it, make sure to use a very low leakage switch to reset the integrator and a very good integration capacitor like NP0 ceramic or a silicon capacitor. I'll write again later to tell you how my integrator solution worked out.
BTW, all in one digital light sensors like those from Hamamatsu and TAOS use an integrator internally.